A Six-Dimensional Taxonomy of Post-Training Adaptation Techniques with Applications in AI Governance

A Six-Dimensional Taxonomy of Post-Training Adaptation Techniques with Applications in AI Governance FARDIN AFDIDEH, Karolinska Institutet, Sweden FERNANDO SEOANE, Karolinska Institutet, Sweden, Karolinska University Hospital, Sweden, University

arXiv:2608.06246v1 [cs.LG] 6 Aug 2026

of Borås, Sweden, and Karolinska University Hospital, Sweden FARHAD ABTAHI, Karolinska Institutet, Sweden, Karolinska University Hospital, Sweden, and KTH Royal Institute of Technology, Sweden Post-training adaptation has become central to modern machine learning practice and includes techniques such as retraining, fine-tuning, parameter-efficient adaptation, alignment, retrieval augmentation, model editing, unlearning, calibration, and Multimodal Instruction Tuning. However, the literature remains fragmented across technique families, model classes and deployment contexts, making it difficult to compare methods or describe how a trained model has been modified. This survey synthesizes the post-training adaptation literature and introduces a six-dimensional taxonomy organized by mechanism, goal, data requirement, persistence, structural scope, and model type. The taxonomy distinguishes commonly conflated terms such as fine-tuning, retrieval augmentation, and prompting, and shows how adaptation strategies evolve from traditional machine learning through deep learning, foundation models, large language models, and multimodal large language models. It also maps relationships among techniques, including inheritance, supersession, hybridization, and layered deployment stacks. The resulting vocabulary can support technical documentation, model-change tracking, and governance analysis. The survey concludes by identifying open challenges in evaluation, reproducibility, persistent inference-time adaptation, unlearning, multimodal adaptation, and governance-aware post-training workflows. Additional Key Words and Phrases: Post-Training Techniques, Model Adaptation Taxonomy, AI Regulatory Compliance.

1

Introduction

Machine Learning (ML) models are rarely static; The environments in which they operate change, e.g., data distributions shift, user behaviors evolve, regulatory requirements change, and new capabilities become desirable. The performance of a model trained on historical data degrades unless the model is adapted to its evolving context. While managing traditional ML drift is a standard operational requirement, the recent proliferation of parameter-efficient and inferencetime adaptation techniques for Foundation Models (FMs) has drastically accelerated the regulatory and change-control challenges surrounding post-deployment modifications. The landscape of posttraining techniques, i.e., methods applied to a model that has already been trained, and that may already be deployed or carry regulatory approval, to modify its behavior, capabilities, or deployment characteristics, is rapidly expanding. The qualifier “post-training” is deliberate, as it identifies the operational context in which compliance obligations regarding modification and change control arise, distinguishing these techniques from the initial training process that precedes deployment. The systematic organization of post-training techniques is a central challenge in both ML system design and AI governance. Authors’ Contact Information: Fardin Afdideh, fardin.afdideh@ki.se, Karolinska Institutet, Department of Clinical Science, Intervention and Technology, Stockholm, Sweden; Fernando Seoane, fernando.seoane@ki.se, Karolinska Institutet, Department of Clinical Science, Intervention and Technology, Stockholm, Sweden and Karolinska University Hospital, Department of Clinical Physiology, Stockholm, Sweden and University of Borås, Department of Textile Technology, Borås, Sweden and Karolinska University Hospital, Department of Medical Technologies, Huddinge, Sweden; Farhad Abtahi, farhad.abtahi@ki.se, Karolinska Institutet, Department of Clinical Science, Intervention and Technology, Stockholm, Sweden and Karolinska University Hospital, Department of Clinical Physiology, Stockholm, Sweden and KTH Royal Institute of Technology, Department of Biomedical Engineering and Health System, Huddinge, Sweden.

2

1.1

Afdideh, Seoane, and Abtahi

Problem and Contributions

At least a dozen techniques exist for modifying ML models, ranging from classical retraining and Transfer Learning (TL) to modern methods such as Parameter-Efficient Fine-Tuning (PEFT), Reinforcement Learning from Human Feedback (RLHF), Reinforcement Learning from Verifiable Rewards (RLVR), and Retrieval-Augmented Generation (RAG). This has produced a fragmented literature characterized by three specific problems: 1) terminological ambiguity, where the same term is used for different concepts, e.g., “Fine-Tuning (FT)” may refer to full parameter updates, partial layer freezing, or adapter training; 2) single-axis taxonomies, which force ambiguous placements (is PEFT an FT technique or an efficiency technique?) and obscure cross-dimensional relationships; and 3) model-type conflation, where current frameworks frequently ignore how model type dictates feasible techniques (Prompt Engineering (PE) is meaningless for a Random Forest; classical Domain Adaptation (DA) has been largely superseded at the Large Language Model (LLM) level by Continued Pre-Training (CPT) and PEFT). These problems can carry legal and financial consequences. For instance, the EU AI Act treats “substantial modification” as a concept that may lead to re-classification of high-risk AI systems and specifies that FT of a third-party model may, depending on context, place the modifier in a provider-like role with associated obligations [3]; however, it provides no technical definition of what constitutes “FT” versus other modifications. The Commission’s General-Purpose AI (GPAI) guidelines introduced a compute-based threshold (one-third of training Floating-Point Operations (FLOPs)), but as Hacker and Holweg argue [79], this proxy may fail to capture low-compute techniques such as Knowledge Editing and Activation Steering that substantially alter behavior with negligible compute [196]. Similarly, the Predetermined Change Control Plan (PCCP) framework of the Food and Drug Administration (FDA) [197] allows pre-specified model modifications; however, audits of historical [215] and current [138] submissions reveal a persistent lack of transparency and omission of technical details required to evaluate model reliability and change. This paper addresses these problems by proposing a six-dimensional characterization taxonomy for ML model adaptation. Rather than organizing techniques into a single hierarchy, the taxonomy assigns every technique a coordinate on six dimensions: 1) Mechanism (D1 ): What changes in the system; 2) Goal (D2 ): Why adapt; 3) Data Requirements (D3 ): What data is needed; 4) Persistence (D4 ): How long the change lasts; 5) Scope (D5 ): How much of the model structure is affected; and 6) Model Type (D6 ): What model is being adapted (Modulating Filter). Dimensions D1 –D5 characterize the declared adaptation intervention through canonical practical profiles; D6 modulates and filters feasible D1 –D5 combinations by the target model and resolves the model-type conflation problem. The taxonomy’s primary artifact is a centerpiece table (Table 8) mapping all 48 techniques to their six-dimensional profiles, complemented by a governance mapping that connects the technical dimensions to regulatory requirements under the National Institute of Standards and Technology (NIST) Artificial Intelligence (AI) Risk Management Framework (RMF), the EU AI Act, the EU Medical Device Regulation (MDR)/In Vitro Diagnostic Regulation (IVDR), and the FDA PCCP. 1.2

Scope

The inclusion principle governing this survey is that a technique is in scope if it modifies the behavior, capabilities, or deployment characteristics of an already-existing ML model, that is, one that has been trained and may already be deployed or carry regulatory approval. The inclusion criterion is technical, but the inclusion of a technique in this taxonomy does not imply that the technique always constitutes a retraining event, a substantial modification, or a reportable device change. A technique is in scope because it can influence the behavior, consistency, reliability, or deployment characteristics of an already-existing model and may therefore need to be described in

A Six-Dimensional Taxonomy of Post-Training Adaptation Techniques with Applications in AI Governance

3

regulated change-control documentation. Whether a specific application of any technique triggers re-certification, a new submission, or a conformity-assessment update is a separate determination that depends on intended purpose, risk profile, clinical or operational impact, and the applicable regulatory pathway. However, the taxonomy does not claim to cover all AI system controls relevant to regulation. Deployment-layer controls, e.g., output guardrails, content moderation filters, watermarking, post-processing rules, and human-in-the-loop review, may also be regulatory-relevant system changes. However, they are excluded here because they do not modify the model’s parameters, inference context, or adaptation pipeline as defined in this taxonomy. Their omission is a scope boundary, not a claim of regulatory irrelevance. The inclusion principle admits two categories of technique. The core category comprises methods applied after training to modify a model’s parameters, inference context, or outputs (e.g., FT, PEFT, alignment, RAG, Knowledge Editing). A boundary extension encompasses training-time data-pipeline strategies, e.g., Data Augmentation, Curriculum Learning, and Active Learning, when applied to an already-existing model as an adaptation mechanism rather than as part of capability building. These boundary-extension techniques are classified as d1 = Pipeline-Mediated with d4 = Pipeline-Dependent in Table 8, making their distinct status visible. Foundational learning paradigms such as supervised, unsupervised, and reinforcement learning, low-level mechanical optimizations, and deployment-layer controls such as output guardrails, watermarking, moderation filters that constrain outputs without modifying model behavior are excluded. The terms post-training techniques and model adaptation refer to the same 48-technique scope but carry context-dependent emphasis. While model adaptation serves as the traditional academic umbrella for modifying a model’s behavior or capabilities, post-training techniques are the contemporary industry and regulatory framing. It foregrounds that the model already exists and may carry compliance obligations such as those under the EU AI Act’s “substantial modification” provisions, the FDA’s PCCP, and MDR post-market monitoring. Consequently, this paper uses “model adaptation” to discuss technical properties and “post-training techniques” to address operational or regulatory change-control contexts. Self-Supervised Learning (SSL) is included only in its adaptation role (e.g., CPT on domainspecific data). The model type hierarchy spans Traditional ML, Deep Learning (DL), FM, LLM, and Multimodal Large Language Model (MLLM). Training is included as a reference row in Table 8, not as a post-training technique, because it establishes a baseline to contrast the inherent properties of post-training techniques.

1.3

Methodology

This study adopts a Multivocal Literature Review (MLR) methodology. This approach is warranted because AI model adaptation is heavily driven by industry practice, with techniques evolving faster than traditional publication cycles. Emergent techniques often appear in framework documentation, developer blogs, and technical reports months or years before they reach formal academic publication. Consequently, restricting this review to peer-reviewed sources alone would omit rapidly evolving concepts that have achieved widespread practitioner adoption but currently lack formal academic treatment. Accordingly, we combined backward and forward snowballing from twelve canonical academic seed clusters (anchored by 18 seed papers) with grey literature from major framework documentation (Hugging Face PEFT, mergekit, EasyEdit, ms-swift), developer reports, and industrial analysis. A candidate received a separate row when it represented an independently selectable adaptation decision unit and had a materially distinct six-dimensional profile from existing rows. Saturation

4

Afdideh, Seoane, and Abtahi

was reached when additional candidates could be represented as variants, sub-techniques, or compositions of existing rows without losing a practically meaningful distinction. Appendix A presents the snowballing chains in thematic rather than integration order. 1.4

Relationship to Existing Surveys

Several authoritative surveys address subsets of the adaptation landscape covered by this taxonomy. Table 1 positions the present work relative to the most relevant existing treatments. The dimensional counts in Table 1 for prior surveys depend on how one parses their taxonomies and reflect this paper’s reading. These approximate counts were coded by the authors using the stated dimension definitions and are intended for positioning rather than exact bibliometric comparison. Table 1. Comparison with existing surveys. Columns indicate techniques covered, classification dimensions, whether model type is a structuring axis, governance mapping, and model-type span. Survey

Tech.

Dims.

D6

Gov.

Model Span

Pan & Yang [153] Zhuang et al. [234] Shi et al. [178] Liu et al. [126] Hospedales et al. [87] Gao et al. [67] Wehner et al. [211] This work

∼8 ∼15 ∼12 ∼10 ∼8 ∼6 ∼10 48

1 1 2 1 3 1 1 6

No Partial Partial No No No No Yes

No No No No No No No Yes

Trad. ML, early DL Trad. ML–DL DL–LLM LLM DL LLM LLM Trad. ML–MLLM

Compared with prior surveys, this taxonomy combines six-dimensional classification, coverage from Traditional ML to MLLMs, and an explicit governance mapping. 1.5

Paper Organization

Section 2 establishes foundational concepts. Section 3 presents the six-dimensional taxonomy with the centerpiece table. Section 3.9 (Appendix C) provides definitions and characterizations of techniques organized by family. Section 4 provides terminology navigation guidance. Section 5 maps taxonomy dimensions to regulatory regimes. Section 6 offers some conclusions, followed by a list of abbreviations and the bibliography. The Supplementary Materials provide the detailed methodology (Appendix A), extended taxonomical category definitions (Appendix B), technique definitions (Appendix C), exploratory structural consistency analysis (Appendix D), and a governance mapping (Appendix E). 2

Foundational Concepts: Training, Drift, and Retraining

This section establishes the baseline adaptation cycle, corresponding to Rows 1 and 8 of the centerpiece table (Table 8). Model training is the process by which an algorithm learns to recognize patterns from data through iterative parameter adjustment to minimize a loss function [147]. The objective is generalization: accurate predictions on new, unseen data, rather than memorization of training examples. Models degrade once deployed because real-world data diverges from the training distribution [65, 129, 171]. The ML community classifies this drift by which property of the joint distribution changes [148]. Covariate shift (or virtual drift) changes the input feature distribution while the underlying relationship between inputs and outputs remains stable [179]; it can often be detected with unsupervised drift monitoring [55, 162]. Concept drift (or real drift) changes the input-output mapping itself, invalidating the model’s learned behavior [213]; remediation typically requires

A Six-Dimensional Taxonomy of Post-Training Adaptation Techniques with Applications in AI Governance

5

labeled data and is costly [129]. At the FM level, degradation often manifests as prompt drift, where user interaction patterns evolve [172, 208], or alignment drift, where the model’s static behavior diverges from shifting safety or cultural standards [48]. Drift detection is fundamentally a governance practice. The NIST AI RMF explicitly treats the failure to monitor for drift as a threat to safety and reliability, making retraining not just a technical optimization, but a necessary compliance control [2, 150]. Model retraining updates a previously trained model on newer or additional data and is a central component of Machine Learning Operations (MLOps) and Continuous Training (CT) methodologies [188, 192]. In the six-dimensional taxonomy, retraining occupies the profile ( d1 = Parametric Update , d2 = Drift Remediation , d3 = Sequential / Incremental , d4 = {Scheduled Permanent, Ad-hoc Permanent} , d5 = Whole-Model , d6 = {ML, DL, FM, LLM, MLLM} ); Its universality across model types reflects its status as the most general-purpose adaptation technique. Triggers include scheduled intervals, performance degradation, drift detection [21, 102, 130, 159], upstream pipeline changes, and manual decisions. Data strategy varies along a stability–adaptability spectrum from offline batch retraining (full corpus, mitigates catastrophic forgetting [29, 136]) through windowed learning to online learning [85] and continual learning [206]. Initial training and retraining share identical d1 = Parametric Update mechanisms but differ strictly on their intended goals ( d2 = Build Capability versus d2 = Drift Remediation ) and data requirements ( d3 = Large Labeled versus d3 = Sequential / Incremental ). The boundary between retraining and continual learning blurs at the incremental end: when retraining uses constrained parameter updates to prevent catastrophic forgetting, it shifts from a d4 = {Scheduled Permanent, Ad-hoc Permanent} , d5 = Whole-Model change to a d4 = {Bounded Cumulative, Unbounded Cumulative} , d5 = Partial one, functionally becoming continual learning (Row 9, Table 8). 3

The Multi-Dimensional Adaptation Taxonomy

This section presents the six-dimensional taxonomy for ML model adaptation. Rather than forcing each adaptation technique into a single slot within a hierarchical tree, the taxonomy defines six conceptually independent axes and positions every technique at a specific coordinate across all six. The dimensions are not empirically orthogonal, since certain pairings are far more common than others, but each answers a distinct question that cannot be derived from the others, and the non-canonical pairings are precisely where the taxonomy provides its greatest analytical value. 3.1

Taxonomy Overview

3.1.1 Motivation. Single-axis taxonomies organize adaptation by one property, e.g., mechanism, domain, or lifecycle stage. They are useful for local comparison but cannot represent techniques that span several functions. For example, PEFT is simultaneously a FT approach, an efficiency strategy, and a model-adaptation method. Single-axis schemes also obscure distinctions among methods with a shared mechanism. PE and RAG both use Context Injection, while their persistence depends on whether a per-request context or a maintained production configuration is being described. They also risk conflating technique properties with target-model properties. FT applied to a ResNet and FT applied to an LLM share a mechanism (D1 ) but may differ in data requirements (D3 ), persistence patterns (D4 ), and complementary adaptation layers. The resulting differences are captured through the remaining dimensions D6 represents the role of model type in constraining practical applicability. The taxonomy addresses these limitations by assigning each practical adaptation technique a six-dimensional profile.

6

Afdideh, Seoane, and Abtahi

3.1.2 The Six Dimensions. Table 2 summarizes five characterizing dimensions and one modulating dimension. D1 –D5 describe the canonical profile of the declared intervention rather than immutable properties of every implementation. D6 describes the target model and constrains practical applicability. This separation reduces confusion between technique properties and model properties. Table 2. The six dimensions: definition and regulatory documentation relevance. D1 –D5 are characterizing dimensions describing canonical properties of the adaptation intervention; D6 is a modulating dimension that filters feasible D1 –D5 combinations. Documentation roles are treated in detail in Section 5. #

Dimension

Core Question

Role

D1

Mechanism

What changes?

D2

Goal

Why adapt?

D3

What data? Data Requirements

D4

Persistence

How long the change lasts?

D5

Scope

How much of the model structure is modified?

D6

Model Type

What model?

Characterizing Describes the technical modification: weights, adapters, retrieval context, prompts, calibration layer, activation path, or data pipeline. Characterizing Links the change to intended purpose and risk rationale: performance maintenance, specialization, alignment, reasoning, knowledge correction, privacy, robustness, or knowledge removal. Characterizing Supports data provenance and evidence traceability: labeled, unlabeled, preference pairs, verifiable task data, external corpus, or forget set. Characterizing May inform temporal change-control analysis: permanent versioning, cumulative monitoring, transient rollbacks, or ephemeral sessions. Characterizing Can support structural modification impact assessments: whole-model, partial, modular/swappable, derived artifacts, or distributed. Defines baseline capability limits and applicability Modulating of GPAI regulatory tiers. Prevents misapplication of model-specific terminology, e.g., LLM-specific terms used for non-LLM systems and vice versa.

Regulatory documentation relevance

3.1.3 How to Read the Taxonomy. Each practical adaptation decision unit is represented by a six-element profile, (d1 , d2 , d3 , d4 , d5 , d6 ), where di denotes the technique’s position on Dimension 𝑖, i.e., di ∈ Di , 𝑖 = 1, . . . , 6. For example, Low-Rank Adaptation (LoRA)-based PEFT is characterized by ( d1 = Parametric Update , d2 = {Task Specialization, Computational Efficiency} , d3 = Small Labeled , d4 = {Scheduled Permanent, Ad-hoc Permanent} , d5 = {Partial, Modular} , d6 = {DL, FM, LLM, MLLM} ). The full set of profiles is compiled in Table 8. Practitioners can enter the taxonomy from any dimension. A deployment engineer constrained by data availability enters through D3 . A compliance officer evaluating governance risk can enter through D2 (Goal) and D4 (Persistence). A researcher comparing techniques might enter through D1 (Mechanism). In all cases, D6 acts as a preliminary filter: the practitioner first identifies their model type, which eliminates infeasible techniques before the remaining dimensions are consulted. Sections 3.2–3.7 define the dimensions, and Table 8 organizes the resulting technique profiles into navigational families. 3.1.4 Navigating Boundary Cases. Techniques marked with § in Table 8 require conditional interpretation. Three recurring cases are recognized, i.e., a technique may span more than one category, may serve both a foundational and an adaptation role, or may remain technically available after being displaced by a more general method. The assigned profile should therefore be interpreted as canonical and should be replaced by the realized implementation when a deployment is documented. Relations among rows are described as umbrella, sub-technique, bridge or hybrid, supersession, and complementarity. Umbrella paradigms contain several independently selectable interventions. Sub-techniques refine a parent row, bridges span families, superseded methods remain technically

A Six-Dimensional Taxonomy of Post-Training Adaptation Techniques with Applications in AI Governance

7

available but are less commonly selected, and complementary techniques may be layered in an ordered deployment stack. A row is retained when it represents an independently selected practical decision unit with a materially distinct profile. Highly experimental cases may fall outside this representation, since the taxonomy prioritizes practical applicability over exhaustive theoretical coverage. 3.2

D1 (Mechanism)

The most fundamental axis is what an adaptation technique modifies. Table 3 organizes the categories, e.g., parameter-updating techniques (full, selective, transient, or distributed updates), structural interventions (architectural modifications), manipulation techniques (parameter composition, transient activation steering, or data pipeline manipulation), and inference-time augmentations (context injection). The d1 = Parametric Update category is the most populated, spanning classical FT, PEFT, prompt learning, alignment (RLHF, Direct Preference Optimization (DPO)), and reasoning-focused Reinforcement Learning (RL) (RLVR). At the LLM level, the taxonomy introduces d1 = Inference-Time Search to formalize test-time compute scaling methods (e.g., OpenAI o1, DeepSeek-R1) and d1 = Activation-Space Manipulation (activation steering), both of which are structurally distinct from gradient-based updates of d1 = Parametric Update (Test-Time Adaptation (TTA)). Across Tables 3–7, statements about the relevance of regulatory documentation identify potential considerations rather than automatic legal consequences; the applicable obligation depends on the actual implementation, intended use, risk profile, and regulatory pathway. Additionally, architectural modification isolates interventions that enact purely topological changes to the computational graph, such as dynamically routing tokens through newly initialized expert sub-networks, expressly excluding updates to pre-existing base weights. From a regulatory standpoint, distinguishing topological shifts from parameter updates may be important for EU AI Act substantial-modification analysis, as introducing new routing logic can alter the software’s structural control flow even if the functional behavior of the base parameters remains unchanged. D1 determines computational cost, reversibility, and infrastructure requirements. As models grow, the distinction between parameter-modifying and non-parameter-modifying methods becomes increasingly consequential: inference-time methods scale independently of model size, while training-time methods scale with it. Consequently, D1 is a key technical axis for assessing whether an adaptation event may implicate compute-based thresholds, such as the EU AI Act’s GPAI guidelines, thereby highlighting the regulatory divergence between parameter-updating methods and inference-time context injection. Certain techniques exhibit composite D1 classifications, e.g., d1 = {Architectural Modification, Parametric Update} for MoE adaptation. It is important to note that algorithmic training protocols, specifically the modification of loss functions, reward shaping, or optimization constraints, e.g., Lagrangian-constrained Proximal Policy Optimization (PPO), are intentionally excluded from D1 categories. While altering an optimization objective is a highly consequential regulatory event, it represents a change to the mathematical protocol calculating the error, rather than the structural mechanism of the modification itself. Regardless of how a safety constraint shapes the gradient, the physical mechanism applied to the model artifact remains a parameter update (e.g., d1 = Parametric Update ). Consequently, shifts in optimization objectives are captured in this taxonomy through shifts in the corresponding D2 (Goal) and D3 (Data Requirements) coordinates, paired with external training protocol documentation. 3.3

D2 (Goal)

This dimension captures the practitioner’s intent, the problem that motivates the choice of adaptation technique. It is the natural entry point for most practitioners, who begin not with “what

8

Afdideh, Seoane, and Abtahi

Table 3. D1 (Mechanism) Categories Category

Definition

Representative Techniques

I. Gradient-Based / Compute-Intensive: Computes gradients against stored weights, consuming training FLOPs. In the context of the EU AI Act’s “substantial modification”, gradient-based methods are especially relevant to compute-threshold analyses. May require or support MDR Annex II-style regression evidence and FDA PCCP pre-specification, depending on device context.

Parametric Update

Executes GD to alter model weights.

Cross-Model Transfer

Training a new model artifact to replicate a teacher model’s outputs or representations.

Training, Retraining, Full/Partial FT, PEFT, MTL, SSL, Prompt Learning, RLHF, DPO, RLVR, CL, Meta-Learning, Calibration KD

II. Gradient-Free / Weight Manipulation: Alters weights or computational graphs without GD. May fall outside compute-based threshold proxies despite causing substantial behavioral or structural changes (e.g., altering control flow or creating compressed artifacts).

Parameter Composition

Algebraic combination of models or mathematical transformation of parameters.

Parametric Compression

Reducing a model’s parametric size, precision, or density to optimize for deployment, without expanding the computational graph. Topological alteration of the computational graph (e.g., adding nodes/routing) expressly without modifying existing base weights.

Architectural Modification

Task Arithmetic, Model Merging, LEACE, Knowledge Editing (closedform) Compression MoE adaptation, dynamic routing

III. Inference-Time / Zero-Footprint: Operates during inference without weight modification and may therefore fall outside weight-update or training-compute thresholds. However, Version-Persistent artifacts (e.g., RAG corpora) may still require system-level transparency documentation under EU AI Act Art. 13 and FDA PCCP.

Context Injection Activation-Space Manipulation Inference-Time Search

Input-context conditioning at inference; zero weight modification Direct algebraic manipulation of intermediate forward-pass activations without weight modification. Allocating variable inference compute through internal generation processes without parameter modification.

PE, RAG, ICL, CE Activation Steering CoT search, Tree of Thoughts, Best-of-𝑛 sampling

IV. Pipeline-Mediated: Adaptation occurs strictly via data pipeline manipulation. Regulatory and change-control documentation is often primarily associated with the downstream training method, although pipeline changes may also require documentation.

Pipeline-Mediated

Characterizes the indirect adaptation pathway for boundary-extension techniques.

Augmentation, Active Learning, Curriculum Learning, Semi-SL

† SSL-as-adaptation (CPT) prescribes its own training objective and modifies parameters; see Table 8, Row 27.

Note: The Pipeline-Mediated category explicitly characterizes the indirect adaptation pathway for boundary-extension techniques, modifying the data pipeline rather than directly altering model parameters.

mechanism should I use?” but with “what problem am I trying to solve?” The goal categories are defined in Table 4. D2 is intentionally practitioner-facing and spans technical, operational, and governance motivations. To support ease of review and governance mapping, the goal categories in Table 4 are organized into five meta-categories, i.e., Capability, Lifecycle, Trust, Knowledge Operations, and Operational Constraints. Regulatory frameworks such as the NIST AI RMF and the EU AI Act express requirements in terms of these high-level goals instead of technical mechanisms. Therefore, D2 can serve as a technical indicator relevant to intended-purpose or intended-use analysis, providing the critical bridge between the technical taxonomy and the regulatory landscape (Section 5). Crucially, the taxonomy formally recognizes Explainability as a standalone adaptation goal. Under frameworks like the EU AI Act Article 13 and the FDA PCCP transparency guidance, developers may face transparency or explainability obligations in high-risk contexts. Post-training modifications

A Six-Dimensional Taxonomy of Post-Training Adaptation Techniques with Applications in AI Governance

9

Table 4. D2 (Goal) Categories Category

Definition

Representative Techniques

I. Capability: Expands or redefines system functionality. Capability-expanding changes may trigger or warrant “substantial modification” reassessments and intended-use review under EU AI Act Annex IV and MDR Annex I. Under FDA PCCP, capability boundaries may need to be pre-specified to avoid unapproved scope expansion. Build Capability Task Specialization

Establish baseline capability from training data. Narrow model expertise to a specific target task.

Capability Extension

Broaden model’s operational scope by adding fundamentally new structural capabilities or modalities. Enhance multi-step problem-solving, planning, or search capabilities. Often verified through programmatic oracles, simulators, or iterative self-competition.

Reasoning

Training FT, PEFT, FSL, MTL, SFT, Active Learning CPT, MoE, Cross-modal alignment, Multimodal tuning RLVR/GRPO

II. Lifecycle: Responds to deployment shifts rather than adding functionality. Lifecycle changes may be relevant to PMS obligations under EU AI Act Art. 72 and MDR Arts. 83–84. Remediation events may need to be logged as part of PMS records; bridging distributional gaps (new populations) may require clinical-evaluation updates, especially where intended population or performance claims change. Drift Remediation Continual Adaptation Distributional Gap Bridging ∗

Restore performance degraded by distribution shift. Accumulate new knowledge without forgetting prior tasks.

Retraining, TTA CL, Incremental Learning

Bridge source-to-target distribution gap (domain shift), same task class.

DA, Active Learning

III. Trust: Relates to EU AI Act Chapter III (Arts. 9–15) requirements for risk, bias, transparency, accuracy, and human oversight. “Trust”-related adaptations may be relevant to conformity-assessment review, even where the technical mechanism appears low-impact. Alignment Safety Robustness Explainability Fairness Reliability

Conform model behavior to human preferences, tone, and helpfulness. Reduce the production of harmful, dangerous, or policy-violating outputs through constitutional-principle enforcement, refusal training, or harm-category filtering. Harden model against adversarial or out-of-distribution inputs. Adapting a model or appending modules to expose internal representations, align reasoning with human logic, or generate post-hoc rationales to satisfy regulatory transparency mandates. Reduce systematic performance or output disparities across demographic or otherwise protected groups (e.g., counterfactual data augmentation, concept erasure). Align expressed model confidence with empirical correctness to support risk-based decision thresholds (e.g., temperature scaling, conformal wrappers).

RLHF, DPO, SFT Safety SFT, RLHF (Refusal), DPO (Refusal), Activation Steering Adversarial Training Probing, Concept Bottlenecks, Attention Alignment LEACE Calibration

IV. Knowledge Operations: The operational locus for GDPR compliance. Knowledge-operation techniques may support workflows relevant to GDPR Art. 17 and data-provenance disclosures. May require updated data-governance documentation under EU AI Act Art. 10. Knowledge Update Remove Knowledge

Inject, update, or correct model knowledge. Selectively erase targeted factual or behavioral knowledge.

RAG, PE, ICL, CE, Knowledge Editing Machine Unlearning

V. Operational Constraints: Alters deployment profile without changing core tasks, potentially implicating compliance considerations such as: GDPR Art. 25 (privacy), GDPR Art. 22 (personalization), AI Act GPAI (compute threshold rules), and AI Act Art. 13 (transparency). May requires distinct configuration-management records. Computational Efficiency Personalization Privacy Preservation Behavior Control

Reduce deployment cost, size, or inference latency. Tailor outputs to individual user preferences and history. Train collaboratively without centralizing sensitive data. Constrain output format, style, persona, or procedural behavior without targeting underlying knowledge or task capability (e.g., system prompts, structured-output steering).

KD, Compression, MoE, Task Arith. / Merging User-embedding personalization, federated personalization, prefix-tuning per-user FL PE

Note: Build Capability represents the initial training phase; it acts as the baseline anchor from which all post-training goals depart. ∗ At the d = {LLM, MLLM} tier, the dedicated DA technique (Row 6) is superseded by CPT and domain-specific PEFT; however, the 6 d2 = Distributional Gap Bridging goal remains achievable at these tiers via SSL/CPT (Row 27).

engineered specifically to satisfy these transparency mandates, e.g., appending probing classifiers, concept bottleneck layers, or attention-alignment fine-tuning, serve an explicit explainability goal that is legally and mechanistically distinct from general d2 = Task Specialization or d2 = Alignment . 3.4

D3 (Data Requirements)

Data availability is frequently the binding constraint in practice; a practitioner with abundant labeled data faces a fundamentally different set of feasible techniques than one with a handful

10

Afdideh, Seoane, and Abtahi

of examples or no training data at all. Crucially, the boundary between a d3 = Small Labeled and a d3 = Large Labeled operates as an order-of-magnitude convention relative to the target model and task complexity, rather than an absolute numerical threshold. Table 5 categorizes these distinct data regimes, ranging from traditional d3 = Small Labeled through specialized inputs like d3 = Paired Multimodal , down to data-free interventions that strictly require d3 = Parameter-Only or d3 = Zero-Shot . Beyond feasibility, D3 helps characterize the regulatory data-provenance burden. Relying on d3 = Preference Pairs introduces annotator-bias governance, e.g., under the EU AI Act; querying an d3 = External Corpus may require copyright, licensing, and provenance controls; and providing a d3 = Forget Set Specification may implicate privacy-compliance obligations, e.g., the GDPR "right to be forgotten". To maintain taxonomic precision, d3 = User Interaction Data is bounded by a strict operational and regulatory definition that isolates it from overlapping categories [3]. Mechanistically, it is restricted to raw, production-phase behavioral exhaust or implicit telemetry directly generated by an end-user during active deployment, e.g., historical query logs, correction clicks, session paths [197]. It is explicitly distinguished from: (i) d3 = Preference Pairs , which requires comparative, explicit human evaluation labels typically gathered in offline curation; (ii) d3 = External Corpus , which constitutes a pre-existing, structurally static reference repository independent of the active user session; and (iii) d3 = Small Labeled , which relies on human annotators adding intentional task labels to raw data. Crucially, this distinction is enforced by its regulatory profile: under the GDPR, d3 = User Interaction Data may carry heightened obligations concerning lawfulness, purpose limitation, and transparency, including the information duties in Articles 13 and 14 that do not apply to curated, non-identifiable training datasets [3]. Similarly, the taxonomy uses operational boundaries to separate overlapping unlabeled data regimes: offline batch samples used to calibrate weights or steering vectors before deployment are classified as d3 = {Unlabeled, Small Labeled} , whereas live queries driving on-the-fly updates are classified as a d3 = Test-Only . The D3 dimension also illustrates the interaction with target models (D6 ): d3 = Paired Multimodal emerges at the d6 = FM level for Representation Foundation Models (RepFMs) (e.g., Contrastive Language-Image Pre-Training (CLIP) [163]), but it becomes a dominant, defining operational constraint at the d6 = MLLM level. Crucially, a shift in D3 can expand or narrow a technique’s operational profile even when structural mechanics are identical. For example, while PE and ICL share identical structural coordinates ( d1 = Context Injection , d5 = Input/Output-Space , d6 = {LLM, MLLM} ), their differing data requirements ( d3 = Zero-Shot versus d3 = Few Demonstrations ) drive distinct variations in their goal and persistence profiles. While both overlap by targeting d2 = Knowledge Update within a d4 = Session-Ephemeral context, PE’s profile broadens to include d2 = Behavior Control and d4 = Version-Persistent artifacts, whereas ICL uniquely introduces d2 = Task Specialization . 3.5 D4 (Persistence) Dimension 4 characterizes the temporal lifespan and operational durability of an adaptation. A central motivation for isolating this dimension is the regulatory and operational requirement for model consistency, defined here as the preservation or controlled modification of model behavior, performance, output characteristics, and intended use, and the maintenance of relevant reliability across versions, updates, deployment contexts, and adaptation layers. Consistency in this sense is not a synonym for stability or accuracy: a model can remain stable under noise while becoming inconsistent across versions and can remain accurate on a benchmark while exhibiting inconsistent behavior on edge cases that matter for regulatory or clinical use. Rather than evaluating structural elements, D4 isolates the state-retention behavior of the model across inference sessions and operational cycles. From a regulatory perspective, this dimension is a primary technical input

A Six-Dimensional Taxonomy of Post-Training Adaptation Techniques with Applications in AI Governance

Table 5. D3 (Data Requirements) Categories Category

Definition

Representative Techniques

I. Human-Annotated Data: Produced via explicit human judgment. Regulatory burden centers on annotator governance, labeling guidelines, and inter-rater reliability. May require documentation of demographic representation and bias mitigation under EU AI Act Art. 10(2)(f–g), plus GDPR Arts. 13/14 disclosures if annotations derive from personal data. Large Labeled Small Labeled

Thousands-to-millions of annotated task examples. Tens-to-hundreds of labeled task examples.

Few Demonstrations Preference Pairs Task Distribution

One-to-five in-context demonstration examples. Human-ranked response pairs encoding behavioral preferences. Distribution of varied tasks for meta-learning episodes or or multi-task training.

Full FT, MTL, Adversarial Training Partial FT, PEFT, gradient-based FSL, Knowledge Editing ICL, metric-based FSL DPO, RLHF Meta-Learning, MTL

II. Synthetic / Programmatic Data: Generated by upstream AI models, simulators, or programmatic rules. Compliance shifts from annotator oversight to provenance tracing of the generating artifact (e.g., teacher biases, constitutional text). Under EU AI Act GPAI rules, teacher-model risks may propagate to the student and should be assessed. Teacher-Derived ConstitutionDerived EnvironmentDerived InstructionResponse Corpus

Data generated by a more capable model acting as an annotator or oracle. Requires auditing the provenance and biases of the upstream teacher model. Data generated iteratively based on explicit textual guidelines or constitutional principles. Requires auditing the constitutional text as a data-governance artifact. Data generated via a physics engine, simulator, play environment, or a model’s own autoregressive self-generation (as an opponent).Requires auditing the simulation parameters and reward mechanics. Heterogeneous prompt-response pairs spanning many tasks, often mixed human-written and synthetic; provenance documentation must cover licensing of scraped sources and the share of model-generated data.

KD CAI, RLAIF Self-Play, Classic RL SFT (Instruction Tuning)

III. Live / Production Data: Generated during active deployment. Depending on whether it involves logged interactions or live inference, this data may implicate GDPR Arts. 6 and 13/14 (consent, lawfulness, purpose limitation) with non-delegable force, or carry no persistent provenance burden but require documented temporal boundaries for on-the-fly adaptations. User Interaction Data Test-Only Sequential / Incremental ∗

Logged user behavior signals collected during deployment. Live, unlabeled test instances encountered on-the-fly during active inference. Time-ordered arriving data batches over a model’s lifecycle.

Personalization methods, implicit feedback loops TTA Retraining, CL

IV. External / Targeted Data: Depending on whether the intervention targets an external knowledge base or specific records flagged for deletion, it either requires copyright, licensing, and quality documentation under EU AI Act Art. 13 transparency rules, or may implicate GDPR Art. 17 (Right to Erasure) analysis and may require verification of genuine record suppression rather than mere obfuscation. External Corpus Forget Set Specification

Queried external knowledge base at inference time. Designated records or facts targeted for erasure.

RAG, CE Machine Unlearning

V. Minimal / Decentralized / Delegated Data: Depending on the specific configuration, these approaches carry varying provenance burdens: they may entail the lowest burden as no new data enters the pipeline, limit centralized traceability, and implicate data-sovereignty and federation rules (e.g., GDPR Ch. V cross-border transfers; HIPAA), or fully delegate regulatory obligations to downstream training methods. Zero-Shot Parameter-Only

Natural language instructions only; no labeled examples. No data; operates exclusively on model parameters.

Decentralized Pipeline-Dependent

Private data distributed across non-collocated clients. The effective data regime is fully inherited from the downstream training method being complemented. Unannotated data used for self-supervised training, continuous adaptation, or distribution sampling for post-training calibration.

Unlabeled †

PE Task Arithmetic, Model Merging, Compression (data-free) FL Data Augmentation, Curriculum Learning, Active Learning SSL, Semi-SL, Compression (calibration), Calibration

VI. Verifier-Defined Ground Truth: Ground truth is established by deterministic programmatic verifiers (e.g., unit tests, math checkers, simulators) rather than human judgment or synthetic generation. This shifts the regulatory burden from annotator-bias tracking to verifier-validity documentation. The verification function acts as a “test oracle” and may require documented validation of the verifier’s soundness and limitations under EU AI Act Art. 15 and FDA/IEC 62304 software-validation standards. Verifiable Task

Tasks with programmatically assessable ground truth.

RLVR/GRPO

VII. Multi-Modal Paired Provenance: Each training example carries independent provenance chains, requiring separate licensing, copyright, and bias documentation for each individual modality. The cross-modal alignment itself should be audited or may need to be assessed for introduced biases. If the pairing involves biometric data used for identification or other special-category data, it may implicate GDPR Art. 9. Furthermore, EU AI Act data-governance provisions may require relevance and representativeness documentation, potentially per modality or pairing, significantly multiplying the compliance burden. Paired Multimodal

Aligned cross-modal pairs (e.g., image-text, audio-text).

Cross-modal alignment, Multimodal instr. tuning

∗ Beyond any applicable GDPR purpose-limitation concerns, each task or domain boundary should generally be logged as a lifecycle event where PMS or post-market monitoring obligations apply. † When the unlabeled sample is drawn from personal health records or user behavioral logs, GDPR Articles 9 (special categories of personal data) and 13/14 obligations apply, elevating the provenance burden irrespective of the absence of explicit labels.

11

12

Afdideh, Seoane, and Abtahi

to version-control and PMS strategy. For example, a modification that alters parameters across multiple deployment intervals may require validation against behavioral drift, whereas transient shifts may require validation of initialization and reset boundaries. To provide clear taxonomic and compliance precision, we categorize the temporal states of adaptation into distinct operational profiles, as detailed in Table 6. This temporal classification can help PCCP authors declare whether an adaptation establishes an enduring structural baseline ( d4 = {Scheduled Permanent, Ad-hoc Permanent} ), updates dynamically across streaming operations ( d4 = {Bounded Cumulative, Unbounded Cumulative} ), or does not necessarily create a persistent parametric modification ( d4 = {Session-Ephemeral, Version-Persistent, Transient} ). The distinction between d4 = Session-Ephemeral and d4 = Transient persistence is structurally profound. d4 = Session-Ephemeral denotes a temporary input context or inference-search state that clears after the request. d4 = Transient denotes temporary internal model state during execution and includes both rolled-back parameter updates, as in TTA, and activation-space interventions, as in Activation Steering. The distinction is therefore temporal and operational rather than purely parametric. Techniques operating on the data pipeline, e.g., Data Augmentation, Active Learning, Curriculum Learning, Semi-SL, are classified as d4 = Pipeline-Dependent . This indicates that their temporal persistence is delegated to, and inherited by, the downstream training method they complement. D4 records the persistence of the realized intervention. Inference-only techniques can still depend on d4 = Version-Persistent system artifacts, so model-parameter persistence and system-level persistence should be documented separately. 3.6

D5 (Scope)

Dimension 5 delineates the structural boundaries and topological localization of the modification within the ML architecture. Dimension 5 isolates where and to what extent the system graph is altered. For compliance officers and engineers executing risk impact assessments (such as under EU MDR Annex II), this dimension determines the structural footprint of a modification, and can inform the scale of regression testing, safety re-validation, and software component separation. The architectural boundaries of post-training interventions are partitioned into six structural meta-categories, as formalized in Table 7. For techniques operating strictly via data-pipeline manipulation, the structural scope is classified as d5 = Pipeline-Dependent , explicitly signaling that the architectural footprint, and the corresponding component component-testing emphasis is primarily inherited from the downstream training method that consumes the data. Crucially, the introduction of the d5 = Input/Output-Space and d5 = Activation-Space categories explicitly isolates techniques (e.g., RAG, PE) that have zero parametric footprint. From a compliance perspective, particularly under MDR Annex II and FDA PCCP impact assessments, these classifications help distinguish systems that may primarily require behavioral validation from those that may also require structural regression testing. The interplay between decoupled structural scope ( d5 = Modular ) and varying persistence levels ( d4 = {Scheduled Permanent, Ad-hoc Permanent} versus d4 = Session-Ephemeral ) has grown rapidly in practical importance with the rise of FMs. Modern deployment patterns frequently layer a permanent, modular PEFT adapter ( d4 = {Scheduled Permanent, Ad-hoc Permanent} , d5 = Modular ) onto a frozen base model, while simultaneously employing ephemeral techniques like PE and RAG ( d4 = {Session-Ephemeral, Version-Persistent} , d5 = Input/Output-Space ) to inject session-specific context through a composite, layered architecture characteristic of the LLM and MLLM tiers. 3.7

Dimension 6 (D6 ): Model Type

Unlike Dimensions D1 –D5 , which characterize inherent properties of an adaptation technique, Dimension D6 characterizes the target model to which a technique is applied. It functions not as

A Six-Dimensional Taxonomy of Post-Training Adaptation Techniques with Applications in AI Governance

13

Table 6. D4 (Persistence) Categories Category

Definition

Representative Techniques

I. Pre-Specifiable Baselines: Scope, cadence, and boundaries can be fully declared pre-deployment in a PCCP or technical documentation. Depending on the intervention, it may declare a fixed temporal cadence, a finite task boundary, or non-parametric system updates—the latter requiring corpus or prompt validation protocols rather than parametric regression testing.

Scheduled Permanent

Pre-planned, periodic parameter updates executed on a fixed temporal or data-volume cadence. Eligible for pre-specification in regulatory plans (e.g., PCCP) as they require predictable, periodic validation rhythms.

Scheduled Retraining, Periodic FT

Bounded Cumulative

Incremental parameter accumulation across a fixed, preenumerable set of new tasks or domains. Enables prespecification of boundaries in regulatory change-control plans (e.g., PCCP).

TIL

Version-

Context-only modification at the model-parameter level, but maintained as a version-controlled, persistent system artifact whose update is a system-level change event.

RAG, CE, Production system prompts, RAG corpora, persistent steering-vector stores, retrieval pipelines

Persistent

II. Event-Driven Unplanned Modification: Triggered by unpredictable events (e.g., sudden drift, emergent vulnerabilities, arbitrary developer decisions). Such modifications may require independent validation, re-clearance review, or conformity-preservation assessment depending on pathway.

Ad-hoc Permanent

Unplanned, event-driven parameter updates triggered by sudden data drift, emergent vulnerabilities, or arbitrary developer decisions. May require event-driven validation and review under the applicable regulatory pathway.

Emergency Patching, Knowledge Editing, Ad-hoc SFT

III. Continuous Surveillance Required: Open-ended, streaming parameter accumulation fundamentally exceeds pre-specifiable PCCP bounds. Such approaches may require enhanced PMS, logging, and behavioral drift monitoring.

Unbounded Cumulative

Open-ended, continuous parameter updates over an infinite or unknown horizon of streaming data. Exceeds pre-specifiable limits, which may motivate enhanced post-market monitoring, logging, and drift controls.

DIL

IV. Session-Scoped: These interventions clear completely and do not persist beyond the inference session. While they may present a lower change-control burden, they may require software validation of reset fidelity to ensure base weights are successfully rolled back and no state carries across sessions.

SessionEphemeral

Context-only modification supplied ad hoc per request or session; state clears post-inference.

User prompts, ICL demonstrations supplied at query time, Test-Time Compute Scaling.

Transient

Temporary inference-session change. Parameter-modifying variants require explicit weight rollback post-inference; activation-modifying variants naturally clear at session end.

TTA, Activation Steering

V. Inherited Burden: Change-control and persistence documentation are entirely delegated to and subsumed by the downstream training event being fed. A separate persistence entry may not be required when the pipeline change is fully covered by the downstream training event; this should be assessed case by case.

PipelineDependent

Persistence fully inherited from complementary downstream training method.

Data Augmentation, Active Learning, Curriculum Learning, SemiSL

an independent characterizing axis but as a modulating dimension: a filter that constrains which combinations of D1 –D5 are feasible for a given model. PE is meaningless for a Random Forest; crossmodal alignment is irrelevant for a text-only LLM. This section defines the model type hierarchy and traces how the available adaptation landscape changes at each level.

14

Afdideh, Seoane, and Abtahi

Table 7. D5 (Scope) Categories Category

Definition

Representative Techniques

I. Full-System Validation: Modifies the entire computational graph or merges models into lineage-opaque artifacts. May warrant full-system regression testing and re-validation, as targeted component rollback is impossible without full checkpoint reversion. May trigger or support the need for stringent provenance and traceability documentation. Whole-Model

All model layers and components modified in full.

Fused Composition

Derived artifact produced by parameter combination of multiple parents, whose behavior is not attributable to any single parent (lineageopaque).

Training, Retraining, Full FT, MTL, Adversarial Training Task Arithmetic, Model Merging, Machine Unlearning (vector negation)

II. Targeted Component Validation: Imposes a partial structural footprint, allowing targeted regression testing of specific components. Rollback feasibility varies by implementation depth. Depending on the intervention, it may be evaluated as a change-control amendment to an existing device, or may require an independent conformity-assessment pathway as a newly created standalone artifact. Partial

Surrogate Model

Compressed Artifact

Only a designated component or layer subset modified. For regulatory rollback documentation, partial modifications vary in reversibility; isolated head updates permit targeted rollback, whereas embedded layer updates require full-model checkpoint restoration. A completely new, independent model artifact trained to replicate a teacher model’s outputs or to provide a learned scoring signal as an intermediate training artifact (e.g., reward model). A surrogate model may need to be assessed as a distinct artifact, depending on its role, intended use, and regulatory pathway. A structurally reduced, pruned, or lower-precision version of the original base model. May be assessed as a modified derivative artifact, depending on its role, intended use, and regulatory pathway.

Partial FT, PEFT, Prompt Learning, Machine Unlearning, Knowledge Editing KD, RLHF

Model Compression

III. Modular Isolation: Modifies the system via decoupled additions while the base architecture remains completely untouched. Modular isolation can provide a comparatively clear rollback documentation pattern (by simply swapping or removing the module). Modular

Decoupled module; base model unchanged and fully rollback-able.

LP, MoE, TIL, Swappable LoRA adapters, Modality-Specific Adapters, Calibration

IV. No Parametric Footprint: Modifies system behavior entirely through input manipulation, output selection, internal search, or transient activation modulation, leaving all internal parameters untouched. Validation may emphasize I/O correctness and behavioral boundaries while retaining integration and end-to-end behavioral checks where appropriate. Input/Output-Space Activation-Space

No modification to any model parameter or internal state; behavioral conditioning occurs exclusively through external input-space manipulation, output selection, or inference search. Direct modulation of intermediate forward-pass representations (e.g., adding learned steering vectors to residual streams) with zero modification to the underlying parameter weights.

PE, RAG, ICL, CE, Test-Time Compute Scaling Activation Steering

V. Distributed Aggregation: Synthesizes global model state from decentralized local updates. Shifts the validation focus from single-system structural regression to cross-client behavioral consistency, differential privacy budget auditing, and aggregation integrity. May implicate crossborder data-transfer regulations where personal data or cross-jurisdictional processing are involved. Distributed

Global model state from federated local-update aggregation.

FL

VI. Inherited Burden: The structural footprint and primary component-validation burden are inherited from the downstream training event. The pipeline intervention should still be documented and assessed where it materially changes the data presented to that event. Pipeline-Dependent

Structural scope is fully inherited from the complementary downstream training method.

Data Augmentation, Active Learning, Curriculum Learning, Semi-SL

3.7.1 The Model Type Hierarchy. The space of ML models forms a tree rather than a linear scale. At the coarsest level, ML divides into Traditional ML (shallow models such as support vector machines, random forests, logistic regression, and k-nearest neighbors) and DL (multi-layer neural networks that learn hierarchical representations, including convolutional, recurrent, and standard Transformer architectures). Within DL, a distinct paradigm has emerged: FMs, defined by Bommasani et al. [31] as large-scale models trained via SSL on broad data and subsequently adapted to a wide range of downstream tasks. While the original definition emphasized breadth of adaptation, FMs in 2024 and beyond are increasingly generative and multimodal.

A Six-Dimensional Taxonomy of Post-Training Adaptation Techniques with Applications in AI Governance

15

At the FM level, the hierarchy branches into two sibling categories: LLMs, which process and generate text and code (e.g., GPT, LLaMA, Mistral), and MLLMs, which process and generate across multiple modalities simultaneously, including text, images, audio, and video (e.g., GPT-4o, Gemini, LLaVA). Structurally, LLMs and MLLMs are siblings under the broader FMs umbrella. While the LLM designation refers text-centric architectures, a model transitions into the MLLM category once it processes additional modalities, such as images or audio. While this taxonomy classifies FMs under a unified structural tier d6 = FM , it is operationally critical to recognize that this node encompasses two distinct architectural lineages: RepFMs and Generative Foundation Models (GenFMs). RepFMs, such as encoder-only architectures (e.g., BERT) and modern embedding models, are optimized for feature extraction, classification, and retrieval. Because they lack decoder blocks, they are structurally incompatible with generation-dependent adaptation mechanisms, such as d1 = Context Injection . Conversely, GenFMs (e.g., T5, BART, and early GPT models) incorporate generative decoders, enabling the use of PE, ICL, and RAG. In this taxonomy, the unified FM node represents the baseline pre-training scale, while the subsequent LLM and MLLM sibling categories act as the direct, specialized descendants of the GenFM lineage. 3.7.2 Model Type as Modulating Dimension. Model type does not simply add an orthogonal axis of classification; it constrains the feasible space of adaptation. Three properties distinguish D6 from D1 –D5 and give the adaptation landscape a relation structure, encompassing inheritance, supersession, and emergent category creation, that cannot be captured by a single hierarchy. First, D6 is a pragmatic applicability and navigation hierarchy rather than a claim that all model classes are mutually exclusive. The hierarchy Traditional ML → DL → FM → {LLM | MLLM} represents common operational applicability: techniques available at a parent node are often, but not invariably, available at descendant nodes. Second, D6 exhibits technique supersession. Moving down the hierarchy not only adds techniques; it can also replace earlier methods with more general alternatives. DA (classical feature alignment) is largely superseded at the LLM level by CPT and domain-specific PEFT. Meta-learning (bi-level optimization for rapid adaptation) is functionally replaced by ICL. Incremental learning is subsumed by continual learning. “Superseded” is not binary: a technique may remain technically available at a higher-level while being practically uncommon because a more general alternative covers the same goal more efficiently. Third, D6 creates new D1 –D5 categories at specific branch points. d1 = Context Injection as a D1 mechanism exists only at the d6 = {LLM, MLLM} level. d3 = Paired Multimodal is most characteristic at the d6 = MLLM level, while related paired-modality regimes may also occur for multimodal representation FMs. d4 = Session-Ephemeral as a D4 category becomes a dominant deployment pattern only at the d6 = {LLM, MLLM} level. D6 does not operate independently of D1 –D5 ; it modulates which categories within each dimension are practically relevant. 3.8

Multi-Dimensional Technique Profiles

Table 8 presents the complete six-dimensional profile of each adaptation technique discussed in this paper. Each row provides the technique’s classification on all six dimensions simultaneously, enabling direct comparison across any axis. The D6 column uses set notation to indicate applicable model types, making inheritance patterns visible: moving down the model hierarchy adds techniques without removing them, except where techniques are superseded by more general alternatives (see Section 3.7.2). The entries are canonical practical profiles rather than universal invariants. Actual deployments should be classified from the realized implementation; To maintain dimensional precision, inclusion in Table 8 is strictly limited to distinct practical adaptation decision units, defined as techniques, strategies, or operational patterns that have a

16

Afdideh, Seoane, and Abtahi

materially distinct six-dimensional profile from existing rows and that a practitioner would consider independently. Consequently, three categories of methods are excluded: (1) Umbrella paradigms that subsume multiple rows, such as TL (which encompasses FT, DA, and FSL; see Appendix C); (2) Pure goals realized through varying mechanisms, such as Personalization, which is instead treated as a D2 category; and (3) Low-level optimizations that constrain the training loop without constituting an adaptation choice in their own right. Table 8. Centerpiece Table: Six-Dimensional Adaptation Technique Profiles. #

Technique

D1 Mechanism

D2 Goal

D3 Data

1

Training

Param. Upd.

Build Capab.

Large Labeled

D4 Persistence

D5 Scope

{Sched. Perm., Whole Ad-hoc Perm.}

D6 Model Type {ML, DL, FM, LLM, MLLM}

I. Knowledge Transfer and Task Specialization 2

FT (full)

Param. Upd.

3

FT (partial)

Param. Upd.

4

PEFT (LoRA, Param. Upd. adapters)

5

ReFT

Param. Upd.

6

DA§

Param. Upd.

7

FSL ¶

Param. Upd.

{Sched. Perm., Ad-hoc Perm.} {Sched. Perm., Task Spec. Small Labeled Ad-hoc Perm.} {Task Spec., {Sched. Perm., Small Labeled Comp. Eff.} Ad-hoc Perm.} {Sched. Perm., Task Spec. Small Labeled Ad-hoc Perm.} Dist. Gap {Unlabeled, {Sched. Perm., Bridging Small Labeled} Ad-hoc Perm.} {Small Labeled, {Sched. Perm., Task Spec. Few Demo.} Ad-hoc Perm.} Task Spec.

Large Labeled

{DL, FM, LLM, MLLM} {DL, FM, LLM, MLLM} Modu- {DL, FM, LLM, MLLM}

Whole Part. {Part., lar} Modular

{LLM, MLLM}

Part.

{DL, FM}

Part.

{DL, FM}

II. Temporal Adaptation and Maintenance 8

Retraining

Param. Upd.

Drift Remed.

Seq./Inc.

9

CL

Param. Upd.

Cont. Adapt.

Seq./Inc.

10

TIL

Param. Upd.

Cont. Adapt.

Seq./Inc.

Param. Upd.

{Cont. Adapt., Seq./Inc. Drift Remed.}

11

DIL

{Sched. Perm., {ML, DL, FM, Whole Ad-hoc Perm.} LLM, MLLM} {Bound. Cu{DL, FM, LLM, mul., Unbound. Part. MLLM} Cumul.} {Part., ModuBound. Cumul. {DL, FM} lar} Unbound. Cu{Part., Whole} {DL, FM} mul.

III. Alignment, Reasoning, and Trustworthiness 12

SFT

Param. Upd.

13

RLHF (LLM Param. Upd. Alignment)∗

14

DPO

{Task Spec., Instr.-Resp. Align., Safety} Corpus {Align., Safety} Pref. Pairs

Param. Upd.

{Align., Safety} Pref. Pairs

{Sched. Perm., Ad-hoc Perm.} {Sched. Perm., Ad-hoc Perm.} {Sched. Perm., Ad-hoc Perm.} {Sched. Perm., Ad-hoc Perm.} {Sched. Perm., Ad-hoc Perm.} {Sched. Perm., Ad-hoc Perm.}

{Part., Whole}

{Part., Whole}

{LLM, MLLM}

{Part., Whole}

{LLM, MLLM}

{Part., Whole}

{LLM, MLLM}

15

RLAIF/CAI

Param. Upd.

Constitution{Align., Safety} Deriv.

16

RLVR/GRPO§

Param. Upd.

Reas.

Verif. Task

17

LP§

Param. Upd.

Expl.

Small Labeled

18

LEACE

Param. Comp.

Fair.

Small Labeled

Ad-hoc Perm.

19

Adversarial Training

Large Labeled

{Sched. Perm., Whole Ad-hoc Perm.}

Param. Upd.

Rob.

{LLM, MLLM}

{Part., Whole, {LLM, MLLM} Surrog. Model}

Modular Part.

{DL, FM, LLM, MLLM} {DL, FM, LLM, MLLM} {DL, FM, LLM}

IV. Training Strategies 20

Meta-Learning Param. Upd.

Task Spec.

Task Dist.

21

MTL

Param. Upd.

Task Spec.

Task Dist.

22

Self-Play‡

Param. Upd.

{Task Spec., EnvironmentAlign., Reas.} Deriv.

{Sched. Perm., Part. Ad-hoc Perm.} {Sched. Perm., Whole Ad-hoc Perm.} {Sched. Perm., Whole Ad-hoc Perm.}

{DL, FM} {DL, FM, LLM, MLLM} {DL, LLM} Continued on next page

A Six-Dimensional Taxonomy of Post-Training Adaptation Techniques with Applications in AI Governance

17

Table 8 – continued from previous page #

Technique

D1 Mechanism

23

Curriculum Learning

Pipe.-Med.

24

Active LearnPipe.-Med. ing

D2 Goal

D3 Data

Task Spec.

Pipe.-Dep.

{Task Spec., Dist. Gap Pipe.-Dep. Bridging}

D4 Persistence

D5 Scope

D6 Model Type

Pipe.-Dep.

Pipe.-Dep.

{ML, DL, FM, LLM, MLLM}

Pipe.-Dep.

Pipe.-Dep.

{ML, DL, FM, LLM, MLLM}

V. Data-Centric and Privacy-Preserving Methods 25

Data AugmenPipe.-Med. tation

{Task Rob.}

26

Semi-SL

Task Spec.

Pipe.-Med.

27

SSL / CPT†

Param. Upd.

28

DP-FT

Param. Upd.

29

FL

Param. Upd.

Spec.,

Pipe.-Dep. {Unlabeled, Small Labeled}

Pipe.-Dep.

Pipe.-Dep.

Pipe.-Dep.

Pipe.-Dep.

{Capab. Ext., Dist. Gap Unlabeled Bridging} {Small Labeled, Priv. Preserv. Large Labeled}

{Unbound. Cumul., Sched. {Part., Whole} Perm.} {Sched. Perm., {Part., Whole} Ad-hoc Perm.}

Priv. Preserv.

Sched. Perm.

Decent.

Dist.

{ML, DL, FM, LLM, MLLM} {ML, DL, FM, LLM, MLLM} {DL, FM, LLM, MLLM} {DL, FM, LLM, MLLM} {ML, DL, FM, LLM, MLLM}

VI. Efficiency and Composition 30 31 32 33 34

Cross-Model Comp. Eff. KD Transf. Model Com- Param. ComComp. Eff. pression pres. Task Arith. & {Capab. Ext., Param. Comp. Model Merging Comp. Eff.} Long-Context {Param. Upd., Capab. Ext. Ext. Arch.} {Param. Upd., {Capab. Ext., MoE§ Arch.} Comp. Eff.}

{Sched. Perm., Ad-hoc Perm.} {Unlabeled, {Sched. Perm., Param.-Only} Ad-hoc Perm.} {Sched. Perm., Param.-Only Ad-hoc Perm.} {Unlabeled, {Sched. Perm., Small Labeled} Ad-hoc Perm.} {Unlabeled, {Sched. Perm., Small Labeled} Ad-hoc Perm.}

Teacher-Deriv.

{DL, FM, LLM, MLLM} Compress. Ar- {DL, FM, LLM, tif. MLLM} {FM, LLM, Fused MLLM} {FM, LLM, Part. MLLM} {Part., Modu- {FM, LLM, lar} MLLM} Surrog. Model

VII. Inference-Time Adaptation 35

PE§

36

Prompt LearnParam. Upd. ing

37

APO

Context Inj.

38

RAG

Context Inj.

39

ICL§

Context Inj.

40

CE

Context Inj.

41

TTA Param. Upd. Test-Time Infer.-Time Compute ScalSearch ing

42

Context Inj.

{Knowl. Upd., Zero-Shot Behav. Ctrl.} Task Spec.

Small Labeled

{Task Spec., Be- {Zero-Shot, hav. Ctrl.} Few Demo.} Knowl. Upd. Ext. Corpus {Task Spec., Few Demo. Knowl. Upd.} {Knowl. Upd., {Ext. Corpus, Person.} User Interact.} Drift Remed. Test-Only Reas.

Zero-Shot

{Sess.-Eph., I/O Space Ver.-Persist.} {Sched. Perm., Part. Ad-hoc Perm.}

{LLM, MLLM} {FM, MLLM}

LLM,

Ver.-Persist.

I/O Space

{LLM, MLLM}

Ver.-Persist.

I/O Space

{LLM, MLLM}

Sess.-Eph.

I/O Space

{LLM, MLLM}

Ver.-Persist.

I/O Space

{LLM, MLLM}

Transient

{Part., Whole}

{DL, FM}

Sess.-Eph.

I/O Space

{LLM, MLLM}

VIII. Calibration, Personalization, and Multimodal Adaptation 43 44 45 46

Calibration

Param. Upd.

Cross-Modal Param. Upd. Alignment Modality-Spec. Param. Upd. Adapters Multimodal InParam. Upd. str. Tuning

{Unlabeled, {Sched. Perm., Small Labeled} Ad-hoc Perm.} {Sched. Perm., Capab. Ext. Pair. Multimod. Ad-hoc Perm.} {Capab. Ext., {Sched. Perm., Pair. Multimod. Comp. Eff.} Ad-hoc Perm.} {Capab. Ext., {Sched. Perm., Pair. Multimod. Align.} Ad-hoc Perm.} Reliab.

Modular

{ML, DL, FM, LLM, MLLM}

Part.

{FM, MLLM}

Modular

MLLM

{Part., Whole}

MLLM

IX. Knowledge Modification and Activation-Based Adaptation 47 48 49

Machine Un- {Param. Upd., {Rem. Knowl., learning Param. Comp.} Priv. Preserv.} Knowledge {Param. Upd., Knowl. Upd. Param. Comp.} Editing§ Activation Act.-Space Ma{Align., Safety} Steering nip.

Forget Spec.

Set

Ad-hoc Perm.

{Part., Fused}

{DL, FM, LLM, MLLM}

Small Labeled

Ad-hoc Perm.

Part.

{LLM, MLLM}

{Unlabeled, Small Labeled}

Transient

Activ.-Space

{LLM, MLLM}

18

Afdideh, Seoane, and Abtahi

Note: Abbreviations are listed in the manuscript before the bibliography. Highly flexible techniques such as CE (e.g., prompting for step-by-step rationales) and RLHF (e.g., penalizing unexplainable refusals) can dynamically adopt d2 = Explainability as a primary goal when engineered specifically to satisfy regulatory transparency mandates. † SSL is classified in its adaptation role (CPT on domain-specific unlabeled data). As a foundational pre-training paradigm, it falls outside the adaptation taxonomy. ‡ Self-play skips the FM tier because standard FM pre-training and FT pipelines do not employ self-play; it re-emerges at the LLM level through RLHF-adjacent techniques (e.g., self-play FT, Self-Play fIne-tuNing (SPIN) [42]) that exploit autoregressive generation as an opponent model. § Classification tensions involving these techniques are discussed in Appendix C. ¶ For in-context FSL (Row 7) at the LLM level, see ICL (Row 39). ∗ While preference-based RL predates foundation models, this row explicitly scopes its modern instantiation for generative alignment at the d6 = {LLM, MLLM} tiers.

Table 8 makes several relationships visible. LLM and MLLM adaptation introduces d1 = Context Injection ,

d1 = Inference-Time Search , d1 = Activation-Space Manipulation , d2 = Alignment , d2 = Reasoning , d2 = Remove Knowledge , and

d4 = Session-Ephemeral persistence. MLLM-specific rows are distinguished by d3 = Paired Multimodal data and MLLM applicability. Machine Unlearning and Knowledge Editing represent different goals, d2 = Remove Knowledge and d2 = Knowledge Update , but both may connect to Task Arithmetic through d1 = Parameter Composition . These relationships illustrate the value of comparing complete dimensional profiles rather than single family labels.

3.9

Technique Definitions

Table 8 organizes the techniques into 9 navigational families: 1) Knowledge Transfer and Task Specialization (Rows 2–7); 2) Temporal Adaptation and Maintenance (Rows 8–11); 3) Alignment, Reasoning, and Trustworthiness (Rows 12–19); 4) Training Strategies (Rows 20– 24); 5) Data-Centric and Privacy-Preserving Methods (Rows 25–29); 6) Efficiency and Composition (Rows 30–34); 7) Inference-Time Adaptation (Rows 35–42); 8) Calibration, Personalization, and Multimodal Adaptation (Rows 43–46); and 9) Knowledge Modification and Activation-Based Adaptation (Rows 47–49). Detailed D1 –D5 category definitions are provided in Appendix B. Detailed technique definitions and boundary cases are provided in Appendix C. 3.10

Exploratory Structural Consistency Analysis

To explore how the six-dimensional characterizations in Table 8 are represented geometrically we computed a Gower-style dissimilarity matrix over D1 –D6 and projected the results using UMAP (full implementation details and geometric mappings are provided in Appendix D). This analysis serves as a diagnostic for whether the proposed functional families occupy coherent regions of the categorical profile space and where boundary cases arise. As illustrated in Appendix D, the projection reveals substantial overlap among the predefined functional families, quantified by a near-zero mean silhouette score. This is expected: the 9 functional families introduced in Section 3.9 (Appendix C) are strictly navigational groupings designed to aid practitioner discovery, not validated statistical clusters. This result supports the need for multi-dimensional characterization: techniques often share one anchor dimension while differing on others. For example, inference-time methods are anchored by ephemeral persistence and context injection, whereas MLLM-specific methods are anchored by paired multimodal data and the MLLM model type. This exploratory representation analysis does not independently validate the taxonomy. 4

Terminology Navigation for Describing Adaptation Pathways

The taxonomy is intended to name and compare adaptation interventions rather than prescribe a preferred method. The workflow below expresses choices in a common vocabulary for research, technical documentation, and governance.

A Six-Dimensional Taxonomy of Post-Training Adaptation Techniques with Applications in AI Governance

4.1

19

Entry Points

Four entry points are commonly used. D6 filters methods by target model, D2 identifies the adaptation purpose, D3 constrains the choice by available data, and D1 with D4 identifies methods compatible with access, compute, and persistence constraints. For example, d2 = Drift Remediation points to retraining or TTA, d2 = Reasoning points to RLVR, d2 = Alignment points to RLHF or DPO, d2 = Knowledge Update points to PE, ICL, RAG, CE, or Knowledge Editing, and d2 = Remove Knowledge points primarily to Machine Unlearning. Data and access constraints then narrow the set further. d3 = Small Labeled data may support PEFT, d3 = Few Demonstrations support ICL, and API-only access generally limits adaptation to d1 = Context Injection or d1 = Inference-Time Search . Activation steering or calibration may require intermediate access, while parameter updates require weight access. 4.2

The Model-Type-First Decision Path

A robust documentation sequence may begin with D6 (Model Type), followed by D2 (Goal) and D3 (Data Requirements), and concludes with D1 (Mechanism), D4 (Persistence), and D5 (Scope). For an LLM, D6 first defines the applicable rows in Table 8. The resulting set is then filtered by the primary adaptation goal and data regime. An LLM intervention for d2 = Knowledge Update , e.g., may be classified as PE with d3 = Zero-Shot instructions, ICL with d3 = Few Demonstrations , or RAG and CE with an d3 = External Corpus . D5 finally records whether the intervention is d5 = Whole-Model , d5 = Partial , d5 = Modular , or confined to the d5 = Input/Output-Space , which supports rollback and change-control documentation. 4.3

Complementary Combinations

Many techniques are complementary rather than mutually exclusive. The ordered composition operator (⊕) denotes a deployment stack in which sequence may influence the resulting behavior and documentation profile. A representative LLM stack is RLVR ( d2 = Reasoning ) ⊕ RLHF ( d2 = Alignment ) ⊕ PEFT ( d2 = Task Specialization ) ⊕ RAG ( d2 = Knowledge Update ) ⊕ PE ( d2 = Behavior Control ) through different mechanisms and persistence profiles. FL ( d2 = Privacy Preservation ) ⊕ PEFT ( d2 = Task Specialization ) provides a second example, in which lightweight adapters are trained across distributed data sources. Privacy claims still depend on safeguards such as secure aggregation, differential privacy, access controls, and an explicit threat model. Each layer should be named and documented separately even when the layers are deployed as one system. 4.4

Common Sources of Terminological and Regulatory Ambiguity

Three ambiguities are particularly important for technical and regulatory documentation. First, FT does not distinguish full FT ( d5 = Whole-Model ), partial FT ( d5 = Partial ), and modular PEFT ( d5 = {Partial, Modular} ). These interventions differ in scope and should be named explicitly. Second, using RAG is incomplete unless the corpus, retrieval method, and update cadence are documented. Third, classical labels such as DA or meta-learning may obscure the actual LLM mechanism, e.g., CPT, domain-specific PEFT, or ICL. D6 helps prevent these terminology mismatches. 5

Governance and Compliance Context

This section does not provide legal advice and does not determine whether a specific model intervention is a substantial modification, a significant change, or a reportable device modification. It shows how the taxonomy can supply the technical vocabulary needed to describe such interventions before regulatory assessment. Determinations of regulatory consequence remain the responsibility

20

Afdideh, Seoane, and Abtahi

of manufacturers, regulators, and notified or auditing bodies acting under the applicable legal framework. As established in Section 3.5, the six-dimensional taxonomy supports the description of changes that affect model consistency, without specifying which changes are acceptable in any particular regulatory context. 5.1

Mapping Taxonomy Dimensions to Regulatory Regimes

Table 2 summarizes the common documentation role of D1 –D6 . The discussion below focuses on how these dimensions are used within the NIST AI RMF, the EU AI Act, the EU MDR and IVDR, and the FDA PCCP. Detailed lifecycle mappings are provided in Appendix E. NIST AI RMF. The seven-stage lifecycle of AI RMF 1.0 (Plan and Design, Collect and Process Data, Build and Use Model, Verify and Validate, Deploy and Use, Operate and Monitor, Use or Impacted By) and the four core functions (Govern, Map, Measure, Manage) identify stage-specific risk-management activities relevant to adaptation [2]; complementary management standards include ISO/IEC 42001 and ISO/IEC 23894. Furthermore, risk management mapping for LLMs should now align with the “Generative AI Profile” (NIST-AI-600-1) [20] and the April 2026 concept note for the “Trustworthy AI Profile” for critical infrastructure [187], which address post-training risks like prompt injection, data poisoning, and model extraction. The “Operate and Monitor” stage invokes d2 = Drift Remediation and d2 = Continual Adaptation goals, which can be mapped to the separate temporal constraints of D4 (Persistence) and structural limits of D5 (Scope) [2]. This isolation maps cleanly to the EU MDR Annex II significant-change protocol, which may require distinct change-control review, depending on whether a modification is structurally localized ( d5 = Partial or d5 = Modular ) and either temporally enduring ( d4 = {Scheduled Permanent, Ad-hoc Permanent} ) or incrementally accumulated ( d4 = {Bounded Cumulative, Unbounded Cumulative} ) [59]. “Verify and Validate” invokes d2 = Alignment , d2 = Reasoning , and d2 = Robustness , including adversarial-threat evaluation across the lifecycle [200]; “Collect and Process Data” invokes D3 , with technique-specific obligations for d3 = Preference Pairs (annotator governance), d3 = Verifiable Task (oracle provenance), and d3 = Decentralized (data sovereignty). EU AI Act. The Act establishes risk-based requirements and post-market obligations that may be informed by the adaptation goal, persistence, and structural scope [3]. The GPAI compute threshold can identify some large FT events, but it may not capture low-compute interventions that materially alter behavior [79]. For example, full FT may consume substantial FLOPs and change the d5 = Whole-Model scope, whereas targeted Knowledge Editing may consume little compute while directly changing a high-stakes factual output. A domain-specific LoRA adapter can also have significant deployment impact despite a small computational footprint. The taxonomy separates compute intensity from purpose, data, persistence, and scope. It also supports documentation of d1 = Context Injection , Trust-related interventions ( d2 = Fairness ), and Fused Composition artifacts ( d5 = Fused Composition ). LEACE, e.g., may provide a specified fairness intervention outside gradientbased compute thresholds, while RAG and production prompts may require documentation when they materially shape a high-risk system. LP can similarly support explainability evidence when internal states are exposed without modifying the validated base model. Furthermore, structural modifications mapped to d5 = Fused Composition create lineage-opaque derived artifacts and may heighten the need for provenance and technical documentation to trace parent-model contributions under Annex IV of the Act. Crucially, this applies equally to models merged via Task Arithmetic (Row 32) and to models subjected to task-vector-negation Machine Unlearning (Row 47), helping avoid inconsistent documentation of mathematically similar operations. EU MDR/IVDR. Under the MDR, AI/ML-driven Medical Device Software (MDSW) is subject to lifecycle obligations beyond initial conformity assessment. Article 83 requires a PMS system proportionate to the risk class [59], Article 84 requires a PMS plan as part of the technical documentation,

A Six-Dimensional Taxonomy of Post-Training Adaptation Techniques with Applications in AI Governance

21

and the IVDR establishes parallel obligations [60]. The taxonomy supports compliance documentation across these obligations: D1 , D4 , and D5 support the technical-documentation requirement (a statement that a model “was updated” is insufficient when MDR Annex II/III-style documentation may call for more precise description, such as “the model received a d1 = Parametric Update producing a d4 = {Scheduled Permanent, Ad-hoc Permanent} , d5 = {Partial, Modular} change via a specified PEFT method”); D2 and D3 support benefit-risk and clinical-evaluation documentation, distinguishing a change driven by d2 = Drift Remediation from one driven by d2 = Task Specialization or d2 = Capability Extension even when d1 and d4 are identical; and D4 with D5 can help assess whether a planned modification may fit within an approved change-control framework, as a substantial modification, or through PMS-driven updates. The MDR’s vigilance reporting obligations (Articles 87–92) and the IVDR’s corresponding obligations (Articles 82–87) are not adaptation events themselves but may trigger them, with the appropriate response depending on the implicated D2 goal. FDA PCCP. The FDA’s PCCP framework [197] allows manufacturers to pre-specify anticipated model modifications and the methodology for implementing them, enabling certain adaptations without new submissions. The broader Good Machine Learning Practice (GMLP) principles, harmonized internationally through the International Medical Device Regulators Forum (IMDRF), emphasize lifecycle data management and performance monitoring that connect to D3 and D4 . Table 9 maps PCCP components to taxonomy dimensions. Table 9. FDA PCCP component ↔ taxonomy dimension PCCP Component

Relevant Taxonomy Dimensions

Description of modification Modification protocol

D1 (what changes), D4 (how lasting), D5 (structural scope), D6 (on what model) D3 (what evidence is required) plus separately stated validation/acceptance criteria D2 (intent), D4 (reach of change), and intended-purpose / safety-effectiveness analysis (outside the framework)

Impact assessment

A PCCP might specify: “The model may be adapted via d1 = Parametric Update (PEFT), for d2 = {Task Specialization, Compu using d3 = Small Labeled (𝑛 ≥ 200), with d4 = {Scheduled Permanent, Ad-hoc Permanent} , d5 = {Partial, Modular} , on a d6 = FM model type.” This specificity is impossible with single-axis descriptions such as “the model may be fine-tuned,” as illustrated by the comparative change-log scenario in Table 10. Table 10. Before/After: Translating a vague change-log entry into a 6D regulatory specification. Example: Medical AI System Update (Change-Log Entry) Before (Standard Prose) Regulatory Ambiguity

“Update v2.1: The model was fine-tuned on recent hospital data to improve performance and incorporate new clinical guidelines.” The term “fine-tuned” fails to specify if the base weights were altered, D5 (Scope), if the change is a transient patch or permanent, D4 (Persistence), or if the guidelines were learned parametrically or retrieved, D1 (Mechanism). This prevents an auditor from determining if a full system re-validation is required. Layer 1 (Parameter Update): PEFT applied to base model. Profile: ( d1 = Parametric Update , d2 = Drift Remediation , d3 = Small Labeled , d4 = Ad-hoc Permanent , d5 = Modular , d6 = LLM ).

After (6D Taxonomy)

Layer 2 (Inference Update): RAG corpus updated with new guidelines. Profile: ( d1 = Context Injection , d2 = Knowledge Update , d3 = External Corpus , d4 = Version-Persistent , d5 = Input/Output-Space , d6 = LLM )

Compliance Payoff

The 6D specification records that the base weights are intended to remain frozen and identifies the adapter and retrieval artifacts as the primary modified components ( d5 = Modular , d5 = Input/Output-Space ).

The taxonomy can support targeted review and rollback planning, but it does not itself prove implementation fidelity or replace end-to-end validation. D4 is particularly relevant to PCCP

22

Afdideh, Seoane, and Abtahi

planning. d4 = Transient , d4 = Version-Persistent , and d4 = {Bounded Cumulative, Unbounded Cumulative} may require different validation, logging, and monitoring arrangements, although the applicable procedure remains case specific. The taxonomy supplies a specification vocabulary rather than an operational governance system. AI/ML Evaluation and Governance Infrastructure for Safety (AEGIS)-type infrastructures [14] may use these profiles as inputs to monitoring and release workflows, but their technical integration remains prospective. 5.2

Why No Single Dimension Determines Regulatory Consequence: Intended Use and Inference-Time Interventions

Inference-time interventions, e.g., RAG, PE, CE, and Activation Steering, may alter deployed behavior without changing stored parameters. Two principles are important. First, a maintained corpus, prompt, index, retriever, or steering artifact can be d4 = Version-Persistent even when the perrequest context is d4 = Session-Ephemeral . Second, regulatory consequence depends on intended use, risk class, and the applicable pathway in addition to the D1 –D6 profile. The same RAG architecture can therefore have different consequences in customer support and clinical decision support. Existing change-control frameworks do not always state clearly how updates to inference-time artifacts should be treated. The taxonomy provides defined intervention types that can be used in future guidance, while it does not resolve the legal classification. Practitioners should therefore document recurring updates to persistent inference-time artifacts even when they are not treated as parameter modifications. A quarterly clinical-guideline corpus update, e.g., is a recurring system-level adaptation. 5.3

Operational Scenario: Medical FM with Layered Adaptation

Consider a hospital deploying a foundation model for radiology report generation across three adaptation layers: Layer 1: Domain-specific PEFT. A LoRA adapter trained on 5,000 institutional reports specializes the base model. Profile: ( d1 = Parametric Update , d2 = Task Specialization , d3 = Small Labeled , d4 = Scheduled Permanent , d5 = {Partial, Modular} , d6 = LLM ). Under the EU AI Act, this layer is likely to be a strong candidate for substantial-modification analysis; under FDA PCCP, it may be appropriate to pre-specify D3 , D4 and D5 constraints. Layer 2: RAG over clinical guidelines. The system retrieves sections from clinical protocols at inference. Profile: ( d1 = Context Injection , d2 = Knowledge Update , d3 = External Corpus , d4 = Version-Persistent , d5 = Input/Output-Space , d6 = LLM ). Although the model’s parameters remain unchanged, its D4 (Persistence) profile is characterized as d4 = Version-Persistent ; therefore, transparency requirements may still apply where retrieval sources materially shape clinical outputs; corpus updates represent a distinct change pathway from parameter updates. Layer 3: PE for output formatting. System prompts enforce structured reporting templates. Profile: ( d1 = Context Injection , d2 = Behavior Control , d3 = Zero-Shot , d4 = Version-Persistent , d5 = Input/Output-Space , d6 = LLM ). This shares d1 and d4 coordinates with Layer 2 but differs on d3 , a distinction invisible in single-axis taxonomies but consequential for change control. The taxonomy makes visible that these three layers operate on different D1 mechanisms, serve different D2 goals, and exhibit different D4 persistence profiles, even though all three contribute to the system’s clinical behavior. A compliance officer can use this decomposition to identify which layers may require separate documentation, separate risk assessment, or separate regulatory review under the applicable pathway. The taxonomy does not determine which procedures apply; it ensures the decomposition needed to answer that question is explicit. Without it, all three layers would be

A Six-Dimensional Taxonomy of Post-Training Adaptation Techniques with Applications in AI Governance

23

conflated under “the model was adapted,” obscuring the distinct dimensional profiles, and therefore the distinct regulatory considerations, of each. 5.4

Trustworthiness and Responsible Adaptation

Beyond specific regulatory regimes, the broader trustworthiness agenda for FM deployments intersects with adaptation at three points that the taxonomy keeps separable. Within the medical domain, this agenda is increasingly formalized by clinical-AI-governance frameworks such as the FUTURE-AI consensus guidelines [112] and AI-specific reporting standards like TRIPOD+AI [44], which emphasize traceability, fairness, and continuous evaluation. Alignment, reasoning, and specialization are distinct D2 goals, not synonyms. A model well-specialized for a clinical domain ( d2 = Task Specialization ) but poorly aligned ( d2 = Alignment ) may produce harmful outputs; a well-aligned model that lacks reasoning depth ( d2 = Reasoning ) may fail at complex cases; a model that reasons well but is unspecialized may be unhelpful. RLHF, DPO, and RLVR target these goals through different mechanisms (D1 ) and data regimes (D3 ), so trustworthiness governance should specify which goal a given intervention addresses rather than collapsing all three under “post-training.” Furthermore, explicitly classifying adaptations under d2 = Fairness —and employing targeted gradientfree interventions like LEACE (Row 18) to apply a mathematically specified projection intended to remove selected protected-attribute concepts from model representations—may support evidence relevant to FDA GMLP principles, MDR clinical-evaluation equity expectations, and EU AI Act bias-mitigation requirements. Meanwhile, d2 = Reliability may help establish risk-based decision thresholds for medical Clinical Decision Support System (CDSS). The clinical regulatory-science community has actively highlighted that static evaluation of such systems is inadequate; models require structured protocols for continual monitoring and updating post-deployment to maintain reliability safely [70, 149]. D3 is the data-provenance and annotator-bias axis. Techniques using Preference Pairs (RLHF, DPO) introduce annotator-selection and labeling-guideline documentation that techniques using only existing labeled data do not [82]. Techniques using d3 = Verifiable Task (RLVR) shift the documentation burden to oracle correctness rather than annotator bias. Techniques using d3 = External Corpus (RAG) require documentation of retrieval-source quality, currency, and licensing. Dynamic benchmarking platforms that evaluate models across multiple D2 goals simultaneously [96] are the operational complement: an adaptation that improves the target D2 goal must not regress on safety-related D2 goals. Assigning an intervention d2 = Remove Knowledge may serve as one technical component of “right to be forgotten” workflows. GDPR and similar regimes may motivate technical workflows involving Machine Unlearning (Row 47) or Knowledge Editing (Row 48), but these methods do not by themselves establish legal compliance. Current methods are approximate rather than provable: benign relearning attacks can reverse unlearning [93], and Rank-One Model Editing (ROME)-style edits degrade after ∼10 sequential applications [89]. Governance frameworks should therefore specify thresholds for knowledge suppression and post-edit validation rather than assume that “remove” is a single operation. 6 6.1

Conclusion and Future Directions Summary of Contributions

This paper proposed a six-dimensional taxonomy for ML model adaptation. The taxonomy decomposes adaptation into Mechanism (D1 ), Goal (D2 ), Data Requirements (D3 ), Persistence (D4 ), Scope (D5 ), and Model Type (D6 ), with D1 –D5 reporting canonical properties of the declared intervention and D6 modulating feasibility by target model. Table 8 classifies 48 techniques on all six dimensions.

24

Afdideh, Seoane, and Abtahi

The governance mapping connects D2 and D4 to potential documentation considerations under the NIST AI RMF, the EU AI Act, the EU MDR/IVDR, and the FDA PCCP. 6.2

Scope Boundaries and Limitations

While the taxonomy is broad, its design incorporates specific boundaries to maintain taxonomic precision: (a) Model Continuum Simplification. The D6 hierarchy discretizes a richer continuum, leaving active-research boundaries (e.g., adversarial training for MLLMs) fluid; (b) Isolation vs. Composition. The taxonomy characterizes individual techniques in isolation, whereas practical deployments layer them in stacks (as modeled in Section 4.3); (c) Classificatory vs. Operational Focus. The taxonomy provides the structural vocabulary but leaves the operational governance of deployment decisions to complementary architectures like AEGIS [14]; (d) MLLM Exhaustiveness. The MLLM-specific rows (44–46) establish the branch structure but defer a comprehensive survey of the rapidly proliferating MLLM landscape to future work; and (e) Empirical Benchmarking. The taxonomy maps technical properties rather than adoption trends or cross-technique performance comparisons; empirical landscape characterization remains a priority for future work. 6.3

Future Directions

Agentic AI and adaptation orchestration. Agentic AI [183] is not a new adaptation technique but an orchestration pattern composing CE, PE, and RAG into a behavioral loop, with per-request execution often operating as d4 = Session-Ephemeral through d1 = Context Injection , while maintained memory and retrieval configurations can be d4 = Version-Persistent . As agents become more capable, the boundary between inference-time orchestration and persistent adaptation may blur: agents that fine-tune their own sub-models, learn user preferences through interaction, or accumulate longterm memory will require the taxonomy to address meta-adaptation, a natural extension of the D2 and D4 axes (e.g., d2 = Learn to Adapt and d4 = Gradually Accumulated ). Additional work is needed on Feature-Training-Inference (FTI) pipeline standardization, privacypreserving combinations of FL and PEFT, and adaptation interfaces for new multimodal data. An access-regime descriptor, i.e., white-box, grey-box, or black-box access, may also be evaluated as a future governance overlay because access constrains both technical feasibility and the modifications that can be operationally controlled. List of Abbreviations ActAdd Activation Addition 52 AdaLoRA Adaptive Low-Rank Adaptation 47, 55 AEGIS AI/ML Evaluation and Governance Infrastructure for Safety 22, 24 AI Artificial Intelligence 2, 5, 8, 20, 24, 43 APO Automatic Prompt Optimization 17, 76 BERT Bidirectional Encoder Representations from Transformers 48, 51 BitFit BIas-Term FIne-Tuning 63 CAA Contrastive Activation Addition 52 CAI Constitutional AI 11, 16, 67 CAVIA Context Adaptation VIA meta-learning 50 CDSS Clinical Decision Support System 23 CE Context Engineering 8, 9, 11, 13, 14, 17, 18, 19, 22, 24, 43, 44, 45, 54, 77 CISPO Clipped IS-weight Policy Optimization 49, 55 CL Continual Learning 8, 9, 11, 16, 43, 50, 64, 65, 84

A Six-Dimensional Taxonomy of Post-Training Adaptation Techniques with Applications in AI Governance

25

CLIP Contrastive Language-Image Pre-Training 10 CoT Chain-of-Thought 8 CPO Contrastive Preference Optimization 49, 55 CPT Continued Pre-Training 2, 3, 8, 9, 15, 17, 18, 19, 63, 72, 76, 80, 81 CPT Context-Aware Prompt Tuning 55 CT Continuous Training 5 DA Domain Adaptation 2, 9, 15, 16, 19, 62, 63, 64, 80 DAPO Decoupled clip and Dynamic sAmpling Policy Optimization 49, 55 DARE Drop And REscale 48, 55, 74 DELLA-Merging Drop and rEscaLe via sampLing with mAgnitude 48 DIL Domain-Incremental Learning 13, 16, 64, 65 DL Deep Learning 3, 4, 14, 15, 65, 69, 70 DoRA Weight-Decomposed Low-Rank Adaptation 47, 55 DP-FT Differentially Private Fine-Tuning 17, 72 DP-PEFT Differentially Private Parameter-Efficient Fine-Tuning 73 DP-SGD Differentially Private Stochastic Gradient Descent 72, 73 DPO Direct Preference Optimization 7, 8, 9, 11, 16, 19, 23, 49, 55, 63, 66, 67, 80, 84, 85 DPR Dense Passage Retrieval 51 DSPy Declarative Self-improving Python 76 DuDe Dual Decomposition of Weights and Singular Value Low-Rank Adaptation 47 DyLoRA Dynamic Low-Rank Adaptation 47 EDA Easy Data Augmentation 53 EMR-Merging Elect, Mask & Rescale-Merging 48 EWC Elastic Weight Consolidation 50 FDA Food and Drug Administration 2, 3, 8, 9, 11, 12, 20, 21, 22, 23, 24 FedPara Federated Parameterization 47 FILA Fisher-Initialization of Low-rank Adapters 52 FL Federated Learning 9, 11, 14, 17, 19, 24, 47, 72, 73, 84, 85, 86 FLOP Floating-Point Operation 2, 8, 20 FM Foundation Model 1, 3, 5, 12, 14, 15, 18, 22, 23, 65, 69, 70, 72, 78, 81, 83, 85 FOMAML First-Order Model Agnostic Meta Learning 50 FSL Few-Shot Learning 9, 11, 16, 18, 50, 62, 63, 64, 80 FT Fine-Tuning 2, 3, 5, 7, 8, 9, 11, 13, 14, 16, 18, 19, 20, 43, 47, 48, 50, 51, 52, 57, 59, 62, 63, 64, 65, 66, 67, 69, 70, 76, 80 FTI Feature-Training-Inference 24 FTL Federated Transfer Learning 48 FV Function Vector 52 GA Gradient Ascent 52 GD Gradient Descent 8, 43, 52, 57, 62, 63, 65, 68, 72, 73, 74, 75, 76 GDPR General Data Protection Regulation 9, 10, 11, 23, 86 GenFM Generative Foundation Model 15, 74 GMLP Good Machine Learning Practice 21, 23 GPAI General-Purpose AI 2, 6, 7, 9, 11, 20 GPT Generative Pre-trained Transformer 51 GPTQ Generative Pre-trained Transformers Quantization 74 GraphRAG Graph Retrieval-Augmented Generation 51 GRPO Group Relative Policy Optimization 9, 11, 16, 44, 45, 49, 54, 55, 67 GSPO Group Sequence Policy Optimization 49, 55

26

Afdideh, Seoane, and Abtahi

HippoRAG Hippocampus-inspired Retrieval-Augmented Generation 51 IA3 Infused Adapter by Inhibiting and Amplifying Inner Activations 47, 55, 63 ICL In-Context Learning 8, 9, 10, 11, 13, 14, 15, 17, 18, 19, 64, 69, 75, 77, 80 IMDRF International Medical Device Regulators Forum 21 IPO Identity Preference Optimization 49, 67 ITI Inference-Time Intervention 52 IVDR In Vitro Diagnostic Regulation 2, 20, 21, 24, 83, 84, 85 KD Knowledge Distillation 8, 9, 11, 14, 17, 73 KL Kullback-Leibler 57 KTO Kahneman-Tversky Optimization 49, 55, 67 LATA Layer-Aware Task Arithmetic 48 LEACE LEAst-squares Concept Erasure 8, 9, 16, 20, 23, 68, 79 LEO Latent Embedding Optimization 50 LLM Large Language Model 2, 3, 4, 5, 6, 7, 12, 13, 15, 16, 18, 19, 20, 47, 48, 49, 51, 52, 53, 54, 63, 64, 65, 66, 69, 70, 72, 75, 76, 77, 80, 81, 85 LOKA Large language mOdel Knowledge updAtes 52 LoKU Low-rank Knowledge Unlearning 52 LoRA Low-Rank Adaptation 6, 14, 16, 20, 22, 47, 55, 63, 75 LOSO Leave One Seed Out 55, 56 LP Linear Probing 14, 16, 20, 67, 68, 81 MAGPRUNE Magnitude-based Pruning 48 MAKE Memory-Associated Knowledge Editing 51 MallowsPO Mallows-model Preference Optimization 49 MAML Model-Agnostic Meta-Learning 50 MC Monte Carlo 55, 56 MDR Medical Device Regulation 2, 3, 8, 9, 12, 20, 21, 23, 24, 83, 84, 85 MDSW Medical Device Software 20, 84 MEMIT Mass-Editing Memory in a Transformer 51, 55, 81 MEND Model Editor Networks with Gradient Decomposition 51, 55, 81 MiLoRA Minor singular component based Low-Rank Adaptation 47 ML Machine Learning 1, 2, 3, 4, 5, 12, 14, 15, 23, 43, 48 MLLM Multimodal Large Language Model 3, 4, 12, 15, 18, 24, 63, 69, 78, 85 MLOps Machine Learning Operations 5 MLP Multi-Layer Perceptron 80 MLR Multivocal Literature Review 3, 43, 46, 54 MoE Mixture of Experts 7, 8, 9, 14, 17, 47, 49, 57, 63, 75, 80 MoELoRA Low-Rank Adaptation as a Mixture of Experts 47 MTL Multi-Task Learning 8, 9, 11, 14, 16, 48, 69, 70, 74 NIST National Institute of Standards and Technology 2, 5, 8, 20, 24, 83, 84 NMT Neural Machine Translation 53 OPRO Optimization by PROmpting 76 ORPO Odds Ratio Preference Optimization 49, 55, 67 PCCP Predetermined Change Control Plan 2, 3, 8, 9, 12, 13, 20, 21, 22, 24 PE Prompt Engineering 2, 5, 8, 9, 10, 11, 12, 13, 14, 15, 17, 19, 22, 24, 44, 54, 75, 76, 77, 78, 80 PEFT Parameter-Efficient Fine-Tuning 2, 3, 5, 6, 7, 8, 9, 11, 12, 14, 15, 16, 19, 21, 22, 24, 43, 44, 47, 54, 55, 62, 63, 65, 66, 68, 69, 72, 75, 78, 80, 84, 85 PEPE Periodic Extrapolation Positional Encodings 74 PI Position Interpolation 74

A Six-Dimensional Taxonomy of Post-Training Adaptation Techniques with Applications in AI Governance

27

PiSSA Principal Singular values and Singular vectors Adaptation 47 PMET Precise Model Editing in a Transformer 51, 55 PMS Post-Market Surveillance 9, 11, 12, 13, 20, 21, 84, 85 PPO Proximal Policy Optimization 7, 49, 55 PRF Probabilistic Relevance Framework 51 PTQ Post-Training Quantization 74 QAT Quantization-Aware Training 74 QLoRA Quantized Low-Rank Adaptation 47, 63 RAFT Retrieval-Augmented Fine-Tuning 51 RAFT Reward rAnked FineTuning 66 RAG Retrieval-Augmented Generation 2, 3, 5, 8, 9, 11, 12, 13, 14, 15, 17, 19, 20, 22, 23, 24, 43, 51, 54, 56, 76, 85 ReFT Representation Fine-Tuning 16, 63 RepE Representation Engineering 43, 52 RepFM Representation Foundation Model 10, 15 RETRO Retrieval-Enhanced Transformer 51 RL Reinforcement Learning 7, 11, 18, 43, 49, 58, 66, 67 RLAIF Reinforcement Learning from AI Feedback 11, 16, 67 RLHF Reinforcement Learning from Human Feedback 2, 7, 8, 9, 11, 14, 16, 18, 19, 23, 49, 54, 63, 66, 67, 80, 81, 84, 85 RLOO REINFORCE Leave-One-Out 49, 55 RLVR Reinforcement Learning from Verifiable Rewards 2, 7, 8, 9, 11, 16, 19, 23, 44, 45, 49, 55, 67, 80, 84, 85 RMF Risk Management Framework 2, 5, 8, 20, 24, 83 RMU Representation Misdirection for Unlearning 52 ROME Rank-One Model Editing 23, 51, 55, 81 RoPE Rotary Positional Embedding 74 rsLoRA rank-stabilized Low-Rank Adaptation 47 SAE Sparse AutoEncoder 52, 80 SAPO Step-Aligned Policy Optimization 49, 55 Self-RAG Self-Reflective Retrieval-Augmented Generation 51 Semi-SL Semi-Supervised Learning 8, 11, 12, 13, 14, 17, 71, 72 SERAC Semi-Parametric Editing with a Retrieval-Augmented Counterfactual Model 51, 55 SFT Supervised Fine-Tuning 9, 11, 13, 16, 49, 55, 65, 66, 72 SGD Stochastic Gradient Descent 48, 50 SimPO Simple Preference Optimization 49, 55, 67 SISA Sharded, Isolated, Sliced, and Aggregated 52 SLERP Spherical Linear Interpolation 48, 54, 55, 74 SPIN Self-Play fIne-tuNing 18 SSL Self-Supervised Learning 3, 8, 9, 11, 14, 17, 18, 72, 81 SVD Singular Value Decomposition 47 SWA Stochastic Weight Averaging 48 TCAV Testing with Concept Activation Vectors 52 TIES-MERGING TrIm, Elect Sign & Merge 48, 55, 74 TIL Task-Incremental Learning 13, 14, 16, 64, 65 TL Transfer Learning 2, 16, 43, 48, 62 TRL Transformer Reinforcement Learning 44, 54, 55 TTA Test-Time Adaptation 7, 9, 11, 12, 13, 17, 19, 77, 80, 81

28

Afdideh, Seoane, and Abtahi

TTT Test-Time Training 77 VeRA Vector-based Random Matrix Adaptation 47, 55 WMDP Weapons of Mass Destruction Proxy 52 YaRN Yet another RoPE extensioN 74 Author Contributions F. Afdideh: Methodology, Software, Validation, Investigation, Data Curation, Formal Analysis, Writing – Original Draft, Writing – Review & Editing, Visualization. F. Seoane: Supervision, Writing – Review & Editing. F. Abtahi: Conceptualization, Methodology, Software, Validation, Investigation, Formal Analysis, Writing – Original Draft, Writing – Review & Editing, Project Administration, Funding Acquisition. Code Availability All project materials, including the source code, are available through the GitHub repository: https://github.com/ki-smile/post-training-taxonomy. Acknowledgments This study was supported by the Stockholm Medical Artificial Intelligence and Learning Environments (SMAILE) core facility at Karolinska Institutet and by EU grant 240037 from EIT Health. The authors used large language models for English proofreading and readability improvements. These tools did not contribute to conceptual content, data analysis, or scientific conclusions. All cited sources were independently verified, and all interpretations remain the sole responsibility of the authors. References [1] [n. d.]. Art. 17 GDPR – Right to erasure (‘right to be forgotten’). https://gdpr-info.eu/art-17-gdpr/ [2] 2023. NIST AI 100-1 – Artificial Intelligence Risk Management Framework (AI RMF 1.0). National Institute of Standards and Technology, U.S. Department of Commerce (Jan. 2023). doi:10.6028/NIST.AI.100-1 [3] 2024. Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 laying down harmonised rules on artificial intelligence and amending Regulations (EC) No 300/2008, (EU) No 167/2013, (EU) No 168/2013, (EU) 2018/858, (EU) 2018/1139 and (EU) 2019/2144 and Directives 2014/90/EU, (EU) 2016/797 and (EU) 2020/1828 (Artificial Intelligence Act) (Text with EEA relevance). http://data.europa.eu/eli/reg/2024/1689/oj [4] 2025. Context Engineering - LLM Memory and Retrieval for AI Agents | Weaviate. https://weaviate.io/blog/contextengineering [5] 2025. Context Engineering Guide: Techniques for AI Agents | LlamaIndex. https://www.llamaindex.ai/blog/contextengineering-what-it-is-and-techniques-to-consider [6] 2025. Context engineering: Why it’s Replacing Prompt Engineering for Enterprise AI Success. https://www.gartner. com/en/articles/context-engineering [7] 2025. Effective context engineering for AI agents. https://www.anthropic.com/engineering/effective-contextengineering-for-ai-agents [8] 2025. Lead the Shift to Context Engineering as Prompt Engineering Fades. https://www.gartner.com/en/documents/ 6781234 [9] 2026. arcee-ai/mergekit. https://github.com/arcee-ai/mergekit [10] 2026. modelscope/ms-swift. original-date: 2023-08-01T15:06:39Z. https://github.com/modelscope/ms-swift [11] 2026. PEFT · Hugging Face. https://huggingface.co/docs/peft/index [12] 2026. TRL - Transformers Reinforcement Learning · Hugging Face. https://huggingface.co/docs/trl/index [13] 2026. zjunlp/EasyEdit. original-date: 2023-05-09T07:48:02Z. https://github.com/zjunlp/EasyEdit [14] Fardin Afdideh, Mehdi Astaraki, Fernando Seoane, and Farhad Abtahi. 2026. AEGIS: An Operational Infrastructure for Post-Market Governance of Adaptive Medical AI Under US and EU Regulatory Frameworks. doi:10.21203/rs.3.rs9199100/v1 [15] Armen Aghajanyan, Sonal Gupta, and Luke Zettlemoyer. 2021. Intrinsic Dimensionality Explains the Effectiveness of Language Model Fine-Tuning. In Proceedings of the 59th Annual Meeting of the Association for Computational

A Six-Dimensional Taxonomy of Post-Training Adaptation Techniques with Applications in AI Governance

29

Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), Chengqing Zong, Fei Xia, Wenjie Li, and Roberto Navigli (Eds.). Association for Computational Linguistics, Online, 7319–7328. doi:10.18653/v1/2021.acl-long.568 [16] Guillaume Alain and Yoshua Bengio. 2018. Understanding intermediate layers using linear classifier probes. doi:10. 48550/arXiv.1610.01644 [17] Andrej Karpathy [@karpathy]. 2025. +1 for "context engineering" over "prompt engineering". https://x.com/karpathy/ status/1937902205765607626 [18] Antreas Antoniou, Harrison Edwards, and Amos Storkey. 2018. How to train your MAML. https://openreview.net/ forum?id=HJGven05Y7 [19] Akari Asai, Zeqiu Wu, Yizhong Wang, Avirup Sil, and Hannaneh Hajishirzi. 2023. Self-RAG: Learning to Retrieve, Generate, and Critique through Self-Reflection. https://openreview.net/forum?id=hSyW5go0v8 [20] Chloe Autio, Reva Schwartz, Jesse Dunietz, Shomik Jain, Martin Stanley, Elham Tabassi, Patrick Hall, and Kamie Roberts. 2024. Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile. National Institute of Standards and Technology (NIST) (July 2024). doi:10.6028/NIST.AI.600-1 [21] Jacob Glenn Ayers, Buvaneswari A. Ramanan, and Manzoor A. Khan. 2025. Detecting Concept Drift in Neural Networks Using Chi-squared Goodness of Fit Testing. doi:10.48550/arXiv.2505.04318 [22] Mohammad Gheshlaghi Azar, Zhaohan Daniel Guo, Bilal Piot, Remi Munos, Mark Rowland, Michal Valko, and Daniele Calandriello. 2024. A General Theoretical Paradigm to Understand Learning from Human Preferences. In Proceedings of The 27th International Conference on Artificial Intelligence and Statistics. PMLR, 4447–4455. https: //proceedings.mlr.press/v238/gheshlaghi-azar24a.html [23] Yuntao Bai, Saurav Kadavath, Sandipan Kundu, Amanda Askell, Jackson Kernion, Andy Jones, Anna Chen, Anna Goldie, Azalia Mirhoseini, Cameron McKinnon, Carol Chen, Catherine Olsson, Christopher Olah, Danny Hernandez, Dawn Drain, Deep Ganguli, Dustin Li, Eli Tran-Johnson, Ethan Perez, Jamie Kerr, Jared Mueller, Jeffrey Ladish, Joshua Landau, Kamal Ndousse, Kamile Lukosuite, Liane Lovitt, Michael Sellitto, Nelson Elhage, Nicholas Schiefer, Noemi Mercado, Nova DasSarma, Robert Lasenby, Robin Larson, Sam Ringer, Scott Johnston, Shauna Kravec, Sheer El Showk, Stanislav Fort, Tamera Lanham, Timothy Telleen-Lawton, Tom Conerly, Tom Henighan, Tristan Hume, Samuel R. Bowman, Zac Hatfield-Dodds, Ben Mann, Dario Amodei, Nicholas Joseph, Sam McCandlish, Tom Brown, and Jared Kaplan. 2022. Constitutional AI: Harmlessness from AI Feedback. doi:10.48550/arXiv.2212.08073 [24] Nora Belrose, David Schneider-Joseph, Shauli Ravfogel, Ryan Cotterell, Edward Raff, and Stella Biderman. 2023. LEACE: Perfect linear concept erasure in closed form. In Advances in Neural Information Processing Systems, A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine (Eds.), Vol. 36. Curran Associates, Inc., 66044–66063. https: //proceedings.neurips.cc/paper_files/paper/2023/file/d066d21c619d0a78c5b557fa3291a8f4-Paper-Conference.pdf [25] Elad Ben Zaken, Yoav Goldberg, and Shauli Ravfogel. 2022. BitFit: Simple Parameter-efficient Fine-tuning for Transformer-based Masked Language-models. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), Smaranda Muresan, Preslav Nakov, and Aline Villavicencio (Eds.). Association for Computational Linguistics, Dublin, Ireland, 1–9. doi:10.18653/v1/2022.acl-short.1 [26] Yoshua Bengio, Jérôme Louradour, Ronan Collobert, and Jason Weston. 2009. Curriculum learning. In Proceedings of the 26th Annual International Conference on Machine Learning (ICML ’09). Association for Computing Machinery, New York, NY, USA, 41–48. doi:10.1145/1553374.1553380 [27] Yoshua Bengio, Mehdi Mirza, Ian Goodfellow, Aaron Courville, and Xia Da. 2014. An Empirical Investigation of Catastrophic Forgeting in Gradient-Based Neural Networks. https://openreview.net/forum?id=oXSw7laxwUpln [28] David Berthelot, Nicholas Carlini, Ian Goodfellow, Nicolas Papernot, Avital Oliver, and Colin A Raffel. 2019. MixMatch: A Holistic Approach to Semi-Supervised Learning. In Advances in Neural Information Processing Systems, Vol. 32. Curran Associates, Inc. https://proceedings.neurips.cc/paper_files/paper/2019/hash/1cd138d0499a68f4bb72bee04bbec2d7Abstract.html [29] Dimitris Bertsimas, Vassilis Digalakis, Yu Ma, and Phevos Paschalidis. 2025. Towards Stable Machine Learning Model Retraining via Slowly Varying Sequences. doi:10.48550/arXiv.2403.19871 [30] Gilles Blanchard, Gyemin Lee, and Clayton Scott. 2011. Generalizing from several related classification tasks to a new unlabeled sample. In Proceedings of the 25th International Conference on Neural Information Processing Systems (NIPS’11). Curran Associates Inc., Red Hook, NY, USA, 2178–2186. [31] Rishi Bommasani, Drew A. Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S. Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, Erik Brynjolfsson, Shyamal Buch, Dallas Card, Rodrigo Castellon, Niladri Chatterji, Annie Chen, Kathleen Creel, Jared Quincy Davis, Dora Demszky, Chris Donahue, Moussa Doumbouya, Esin Durmus, Stefano Ermon, John Etchemendy, Kawin Ethayarajh, Li Fei-Fei, Chelsea Finn, Trevor Gale, Lauren Gillespie, Karan Goel, Noah Goodman, Shelby Grossman, Neel Guha, Tatsunori Hashimoto, Peter Henderson, John Hewitt, Daniel E. Ho, Jenny Hong, Kyle Hsu, Jing Huang, Thomas Icard, Saahil Jain, Dan Jurafsky, Pratyusha Kalluri, Siddharth Karamcheti, Geoff Keeling, Fereshte Khani, Omar Khattab, Pang Wei Koh, Mark Krass, Ranjay

30

Afdideh, Seoane, and Abtahi

Krishna, Rohith Kuditipudi, Ananya Kumar, Faisal Ladhak, Mina Lee, Tony Lee, Jure Leskovec, Isabelle Levent, Xiang Lisa Li, Xuechen Li, Tengyu Ma, Ali Malik, Christopher D. Manning, Suvir Mirchandani, Eric Mitchell, Zanele Munyikwa, Suraj Nair, Avanika Narayan, Deepak Narayanan, Ben Newman, Allen Nie, Juan Carlos Niebles, Hamed Nilforoshan, Julian Nyarko, Giray Ogut, Laurel Orr, Isabel Papadimitriou, Joon Sung Park, Chris Piech, Eva Portelance, Christopher Potts, Aditi Raghunathan, Rob Reich, Hongyu Ren, Frieda Rong, Yusuf Roohani, Camilo Ruiz, Jack Ryan, Christopher Ré, Dorsa Sadigh, Shiori Sagawa, Keshav Santhanam, Andy Shih, Krishnan Srinivasan, Alex Tamkin, Rohan Taori, Armin W. Thomas, Florian Tramèr, Rose E. Wang, William Wang, Bohan Wu, Jiajun Wu, Yuhuai Wu, Sang Michael Xie, Michihiro Yasunaga, Jiaxuan You, Matei Zaharia, Michael Zhang, Tianyi Zhang, Xikun Zhang, Yuhui Zhang, Lucia Zheng, Kaitlyn Zhou, and Percy Liang. 2022. On the Opportunities and Risks of Foundation Models. doi:10.48550/arXiv.2108.07258 [32] Sebastian Borgeaud, Arthur Mensch, Jordan Hoffmann, Trevor Cai, Eliza Rutherford, Katie Millican, George Bm Van Den Driessche, Jean-Baptiste Lespiau, Bogdan Damoc, Aidan Clark, Diego De Las Casas, Aurelia Guy, Jacob Menick, Roman Ring, Tom Hennigan, Saffron Huang, Loren Maggiore, Chris Jones, Albin Cassirer, Andy Brock, Michela Paganini, Geoffrey Irving, Oriol Vinyals, Simon Osindero, Karen Simonyan, Jack Rae, Erich Elsen, and Laurent Sifre. 2022. Improving Language Models by Retrieving from Trillions of Tokens. In Proceedings of the 39th International Conference on Machine Learning. PMLR, 2206–2240. https://proceedings.mlr.press/v162/borgeaud22a.html [33] Lucas Bourtoule, Varun Chandrasekaran, Christopher A. Choquette-Choo, Hengrui Jia, Adelin Travers, Baiwu Zhang, David Lie, and Nicolas Papernot. 2021. Machine Unlearning. In 2021 IEEE Symposium on Security and Privacy (SP). 141–159. ISSN: 2375-1207. doi:10.1109/SP40001.2021.00019 [34] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language models are few-shot learners. In Proceedings of the 34th International Conference on Neural Information Processing Systems (NIPS ’20). Curran Associates Inc., Red Hook, NY, USA, 1877–1901. [35] Eric L. Buehler and Markus J. Buehler. 2024. X-LoRA: Mixture of low-rank adapter experts, a flexible framework for large language models with applications in protein mechanics and molecular design. APL Mach. Learn. 2, 2 (May 2024), 026119. doi:10.1063/5.0203126 [36] Yinzhi Cao and Junfeng Yang. 2015. Towards Making Systems Forget with Machine Unlearning. In 2015 IEEE Symposium on Security and Privacy. 463–480. doi:10.1109/SP.2015.35 [37] Rich Caruana. 1997. Multitask Learning. Machine Learning 28, 1 (July 1997), 41–75. doi:10.1023/A:1007379606734 [38] Sungmin Cha, Sungjun Cho, Dasol Hwang, and Moontae Lee. 2024. Towards Robust and Parameter-Efficient Knowledge Unlearning for LLMs. https://openreview.net/forum?id=1ExfUpmIW4 [39] Haoxian Chen, Hanyang Zhao, Henry Lam, David Yao, and Wenpin Tang. 2024. MallowsPO: Fine-Tune Your LLM with Preference Dispersions. https://openreview.net/forum?id=d8cnezVcaW [40] Shouyuan Chen, Sherman Wong, Liangjian Chen, and Yuandong Tian. 2023. Extending Context Window of Large Language Models via Positional Interpolation. doi:10.48550/arXiv.2306.15595 [41] Yan-Lun Chen, Yi-Ru Wei, Chia-Yi Hsu, Chia-Mu Yu, Chun-Ying Huang, Ying-Dar Lin, Yu-Sung Wu, and Wei-Bin Lee. 2025. Layer-Aware Task Arithmetic: Disentangling Task-Specific and Instruction-Following Knowledge. In Findings of the Association for Computational Linguistics: EMNLP 2025, Christos Christodoulopoulos, Tanmoy Chakraborty, Carolyn Rose, and Violet Peng (Eds.). Association for Computational Linguistics, Suzhou, China, 12033–12054. doi:10.18653/v1/2025.findings-emnlp.644 [42] Zixiang Chen, Yihe Deng, Huizhuo Yuan, Kaixuan Ji, and Quanquan Gu. 2024. Self-play fine-tuning converts weak language models to strong language models. In Proceedings of the 41st International Conference on Machine Learning (ICML’24, Vol. 235). JMLR.org, Vienna, Austria, 6621–6642. [43] Paul F. Christiano, Jan Leike, Tom B. Brown, Miljan Martic, Shane Legg, and Dario Amodei. 2017. Deep reinforcement learning from human preferences. In Proceedings of the 31st International Conference on Neural Information Processing Systems (NIPS’17). Curran Associates Inc., Red Hook, NY, USA, 4302–4310. https://dl.acm.org/doi/10.5555/3294996. 3295184 [44] Gary S. Collins, Karel G. M. Moons, Paula Dhiman, Richard D. Riley, Andrew L. Beam, Ben Van Calster, Marzyeh Ghassemi, Xiaoxuan Liu, Johannes B. Reitsma, Maarten van Smeden, Anne-Laure Boulesteix, Jennifer Catherine Camaradou, Leo Anthony Celi, Spiros Denaxas, Alastair K. Denniston, Ben Glocker, Robert M. Golub, Hugh Harvey, Georg Heinze, Michael M. Hoffman, André Pascal Kengne, Emily Lam, Naomi Lee, Elizabeth W. Loder, Lena MaierHein, Bilal A. Mateen, Melissa D. McCradden, Lauren Oakden-Rayner, Johan Ordish, Richard Parnell, Sherri Rose, Karandeep Singh, Laure Wynants, and Patricia Logullo. 2024. TRIPOD+AI statement: updated guidance for reporting clinical prediction models that use regression or machine learning methods. BMJ 385 (April 2024), e078378. doi:10. 1136/bmj-2023-078378

A Six-Dimensional Taxonomy of Post-Training Adaptation Techniques with Applications in AI Governance

31

[45] Ekin D. Cubuk, Barret Zoph, Dandelion Mané, Vijay Vasudevan, and Quoc V. Le. 2019. AutoAugment: Learning Augmentation Strategies From Data. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 113–123. ISSN: 2575-7075. doi:10.1109/CVPR.2019.00020 [46] Hoagy Cunningham, Aidan Ewart, Logan Riggs, Robert Huben, and Lee Sharkey. 2023. Sparse Autoencoders Find Highly Interpretable Features in Language Models. doi:10.48550/arXiv.2309.08600 [47] Damai Dai, Li Dong, Yaru Hao, Zhifang Sui, Baobao Chang, and Furu Wei. 2022. Knowledge Neurons in Pretrained Transformers. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), Smaranda Muresan, Preslav Nakov, and Aline Villavicencio (Eds.). Association for Computational Linguistics, Dublin, Ireland, 8493–8502. doi:10.18653/v1/2022.acl-long.581 [48] Amitava Das, Vinija Jain, and Aman Chadha. 2025. TRACEALIGN - Tracing the Drift: Attributing Alignment Failures to Training-Time Belief Sources in LLMs. (Oct. 2025). https://openreview.net/forum?id=6mzS7NAiui [49] MohammadReza Davari and Eugene Belilovsky. 2025. Model Breadcrumbs: Scaling Multi-task Model Merging with Sparse Masks. In Computer Vision – ECCV 2024, Aleš Leonardis, Elisa Ricci, Stefan Roth, Olga Russakovsky, Torsten Sattler, and Gül Varol (Eds.). Springer Nature Switzerland, Cham, 270–287. doi:10.1007/978-3-031-73226-3_16 [50] Nicola De Cao, Wilker Aziz, and Ivan Titov. 2021. Editing Factual Knowledge in Language Models. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, Marie-Francine Moens, Xuanjing Huang, Lucia Specia, and Scott Wen-tau Yih (Eds.). Association for Computational Linguistics, Online and Punta Cana, Dominican Republic, 6491–6506. doi:10.18653/v1/2021.emnlp-main.522 [51] Pala Tej Deep, Rishabh Bhardwaj, and Soujanya Poria. 2024. DELLA-Merging: Reducing Interference in Model Merging through Magnitude-Based Sampling. doi:10.48550/arXiv.2406.11617 [52] Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. 2023. QLoRA: Efficient Finetuning of Quantized LLMs. In Advances in Neural Information Processing Systems, A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine (Eds.), Vol. 36. Curran Associates, Inc., 10088–10115. https://proceedings.neurips.cc/paper_ files/paper/2023/file/1feb87871436031bdc0f2beaa62a049b-Paper-Conference.pdf [53] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. doi:10.48550/arXiv.1810.04805 [54] Hanze Dong, Wei Xiong, Deepanshu Goyal, Yihan Zhang, Winnie Chow, Rui Pan, Shizhe Diao, Jipeng Zhang, Kashun Shum, and Tong Zhang. 2023. RAFT: Reward rAnked FineTuning for Generative Foundation Model Alignment. doi:10.48550/arXiv.2304.06767 [55] Denis Moreira dos Reis, Peter Flach, Stan Matwin, and Gustavo Batista. 2016. Fast Unsupervised Online Drift Detection Using Incremental Kolmogorov-Smirnov Test. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD ’16). Association for Computing Machinery, New York, NY, USA, 1545–1554. doi:10.1145/2939672.2939836 [56] Shihan Dou, Enyu Zhou, Yan Liu, Songyang Gao, Wei Shen, Limao Xiong, Yuhao Zhou, Xiao Wang, Zhiheng Xi, Xiaoran Fan, Shiliang Pu, Jiang Zhu, Rui Zheng, Tao Gui, Qi Zhang, and Xuanjing Huang. 2024. LoRAMoE: Alleviating World Knowledge Forgetting in Large Language Models via MoE-Style Plugin. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), Lun-Wei Ku, Andre Martins, and Vivek Srikumar (Eds.). Association for Computational Linguistics, Bangkok, Thailand, 1932–1945. doi:10.18653/v1/2024.acl-long.106 [57] Darren Edge, Ha Trinh, Newman Cheng, Joshua Bradley, Alex Chao, Apurva Mody, Steven Truitt, Dasha Metropolitansky, Robert Osazuwa Ness, and Jonathan Larson. 2025. From Local to Global: A Graph RAG Approach to Query-Focused Summarization. doi:10.48550/arXiv.2404.16130 [58] Kawin Ethayarajh, Winnie Xu, Niklas Muennighoff, Dan Jurafsky, and Douwe Kiela. 2024. Model alignment as prospect theoretic optimization. In Proceedings of the 41st International Conference on Machine Learning (ICML’24, Vol. 235). JMLR.org, Vienna, Austria, 12634–12651. [59] European Parliament and of the Council. 2017. Regulation (EU) 2017/745 of the European Parliament and of the Council of 5 April 2017 on medical devices, amending Directive 2001/83/EC, Regulation (EC) No 178/2002 and Regulation (EC) No 1223/2009 and repealing Council Directives 90/385/EEC and 93/42/EEC (Text with EEA relevance. ). http://data.europa.eu/eli/reg/2017/745/oj [60] European Parliament and of the Council. 2017. Regulation (EU) 2017/746 of the European Parliament and of the Council of 5 April 2017 on in vitro diagnostic medical devices and repealing Directive 98/79/EC and Commission Decision 2010/227/EU (Text with EEA relevance. ). http://data.europa.eu/eli/reg/2017/746/oj [61] Eoin Farrell, Yeu-Tong Lau, and Arthur Conmy. 2024. Applying Sparse Autoencoders to Unlearn Knowledge in Language Models. (Oct. 2024). https://openreview.net/forum?id=ZtvRqm6oBu [62] Steven Y. Feng, Varun Gangal, Jason Wei, Sarath Chandar, Soroush Vosoughi, Teruko Mitamura, and Eduard Hovy. 2021. A Survey of Data Augmentation Approaches for NLP. In Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, Chengqing Zong, Fei Xia, Wenjie Li, and Roberto Navigli (Eds.). Association for Computational

32

Afdideh, Seoane, and Abtahi

Linguistics, Online, 968–988. doi:10.18653/v1/2021.findings-acl.84 [63] Chelsea Finn, Pieter Abbeel, and Sergey Levine. 2017. Model-agnostic meta-learning for fast adaptation of deep networks. In Proceedings of the 34th International Conference on Machine Learning - Volume 70 (ICML’17). JMLR.org, Sydney, NSW, Australia, 1126–1135. https://dl.acm.org/doi/10.5555/3305381.3305498 [64] Elias Frantar, Saleh Ashkboos, Torsten Hoefler, and Dan Alistarh. 2023. GPTQ: Accurate Post-Training Quantization for Generative Pre-trained Transformers. doi:10.48550/arXiv.2210.17323 [65] João Gama, Indrė Žliobaitė, Albert Bifet, Mykola Pechenizkiy, and Abdelhamid Bouchachia. 2014. A survey on concept drift adaptation. ACM Comput. Surv. 46, 4 (March 2014), 44:1–44:37. doi:10.1145/2523813 [66] Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario Marchand, and Victor Lempitsky. 2016. Domain-adversarial training of neural networks. J. Mach. Learn. Res. 17, 1 (Jan. 2016), 2096–2030. https://dl.acm.org/doi/10.5555/2946645.2946704 [67] Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yi Dai, Jiawei Sun, Meng Wang, and Haofen Wang. 2024. Retrieval-Augmented Generation for Large Language Models: A Survey. doi:10.48550/arXiv.2312.10997 [68] Vahid Garousi, Michael Felderer, and Mika V. Mäntylä. 2019. Guidelines for including grey literature and conducting multivocal literature reviews in software engineering. Information and Software Technology 106 (Feb. 2019), 101–121. doi:10.1016/j.infsof.2018.09.006 [69] Mor Geva, Roei Schuster, Jonathan Berant, and Omer Levy. 2021. Transformer Feed-Forward Layers Are Key-Value Memories. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, Marie-Francine Moens, Xuanjing Huang, Lucia Specia, and Scott Wen-tau Yih (Eds.). Association for Computational Linguistics, Online and Punta Cana, Dominican Republic, 5484–5495. doi:10.18653/v1/2021.emnlp-main.446 [70] Stephen Gilbert, Matthew Fenech, Martin Hirsch, Shubhanan Upadhyay, Andrea Biasiucci, and Johannes Starlinger. 2021. Algorithm Change Protocols in the Regulation of Adaptive Machine Learning–Based Medical Devices. J Med Internet Res 23, 10 (Oct. 2021), e30545. doi:10.2196/30545 [71] Antonio Ginart, Melody Guan, Gregory Valiant, and James Zou. 2019. Making AI Forget You: Data Deletion in Machine Learning. In Advances in Neural Information Processing Systems, Vol. 32. Curran Associates, Inc. https: //proceedings.neurips.cc/paper/2019/hash/cb79f8fa58b91d3af6c9c991f63962d3-Abstract.html [72] Jianping Gou, Baosheng Yu, Stephen J. Maybank, and Dacheng Tao. 2021. Knowledge Distillation: A Survey. Int J Comput Vis 129, 6 (June 2021), 1789–1819. doi:10.1007/s11263-021-01453-z [73] Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q. Weinberger. 2017. On calibration of modern neural networks. In Proceedings of the 34th International Conference on Machine Learning - Volume 70 (ICML’17). JMLR.org, Sydney, NSW, Australia, 1321–1330. https://dl.acm.org/doi/10.5555/3305381.3305518 [74] Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Peiyi Wang, Qihao Zhu, Runxin Xu, Ruoyu Zhang, Shirong Ma, Xiao Bi, Xiaokang Zhang, Xingkai Yu, Yu Wu, Z. F. Wu, Zhibin Gou, Zhihong Shao, Zhuoshu Li, Ziyi Gao, Aixin Liu, Bing Xue, Bingxuan Wang, Bochao Wu, Bei Feng, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chong Ruan, Damai Dai, Deli Chen, Dongjie Ji, Erhang Li, Fangyun Lin, Fucong Dai, Fuli Luo, Guangbo Hao, Guanting Chen, Guowei Li, H. Zhang, Hanwei Xu, Honghui Ding, Huazuo Gao, Hui Qu, Hui Li, Jianzhong Guo, Jiashi Li, Jingchang Chen, Jingyang Yuan, Jinhao Tu, Junjie Qiu, Junlong Li, J. L. Cai, Jiaqi Ni, Jian Liang, Jin Chen, Kai Dong, Kai Hu, Kaichao You, Kaige Gao, Kang Guan, Kexin Huang, Kuai Yu, Lean Wang, Lecong Zhang, Liang Zhao, Litong Wang, Liyue Zhang, Lei Xu, Leyi Xia, Mingchuan Zhang, Minghua Zhang, Minghui Tang, Mingxu Zhou, Meng Li, Miaojun Wang, Mingming Li, Ning Tian, Panpan Huang, Peng Zhang, Qiancheng Wang, Qinyu Chen, Qiushi Du, Ruiqi Ge, Ruisong Zhang, Ruizhe Pan, Runji Wang, R. J. Chen, R. L. Jin, Ruyi Chen, Shanghao Lu, Shangyan Zhou, Shanhuang Chen, Shengfeng Ye, Shiyu Wang, Shuiping Yu, Shunfeng Zhou, Shuting Pan, S. S. Li, Shuang Zhou, Shaoqing Wu, Tao Yun, Tian Pei, Tianyu Sun, T. Wang, Wangding Zeng, Wen Liu, Wenfeng Liang, Wenjun Gao, Wenqin Yu, Wentao Zhang, W. L. Xiao, Wei An, Xiaodong Liu, Xiaohan Wang, Xiaokang Chen, Xiaotao Nie, Xin Cheng, Xin Liu, Xin Xie, Xingchao Liu, Xinyu Yang, Xinyuan Li, Xuecheng Su, Xuheng Lin, X. Q. Li, Xiangyue Jin, Xiaojin Shen, Xiaosha Chen, Xiaowen Sun, Xiaoxiang Wang, Xinnan Song, Xinyi Zhou, Xianzu Wang, Xinxia Shan, Y. K. Li, Y. Q. Wang, Y. X. Wei, Yang Zhang, Yanhong Xu, Yao Li, Yao Zhao, Yaofeng Sun, Yaohui Wang, Yi Yu, Yichao Zhang, Yifan Shi, Yiliang Xiong, Ying He, Yishi Piao, Yisong Wang, Yixuan Tan, Yiyang Ma, Yiyuan Liu, Yongqiang Guo, Yuan Ou, Yuduan Wang, Yue Gong, Yuheng Zou, Yujia He, Yunfan Xiong, Yuxiang Luo, Yuxiang You, Yuxuan Liu, Yuyang Zhou, Y. X. Zhu, Yanping Huang, Yaohui Li, Yi Zheng, Yuchen Zhu, Yunxian Ma, Ying Tang, Yukun Zha, Yuting Yan, Z. Z. Ren, Zehui Ren, Zhangli Sha, Zhe Fu, Zhean Xu, Zhenda Xie, Zhengyan Zhang, Zhewen Hao, Zhicheng Ma, Zhigang Yan, Zhiyu Wu, Zihui Gu, Zijia Zhu, Zijun Liu, Zilin Li, Ziwei Xie, Ziyang Song, Zizheng Pan, Zhen Huang, Zhipeng Xu, Zhongyu Zhang, and Zhen Zhang. 2025. DeepSeek-R1 incentivizes reasoning in LLMs through reinforcement learning. Nature 645, 8081 (Sept. 2025), 633–638. doi:10.1038/s41586-025-09422-z [75] Wei Guo, Fuzhen Zhuang, Xiao Zhang, Yiqi Tong, and Jin Dong. 2024. A comprehensive survey of federated transfer learning: challenges, methods and applications. Front. Comput. Sci. 18, 6 (July 2024), 186356. doi:10.1007/s11704-02440065-x

A Six-Dimensional Taxonomy of Post-Training Adaptation Techniques with Applications in AI Governance

33

[76] Akshat Gupta, Anurag Rao, and Gopala Anumanchipalli. 2024. Model Editing at Scale leads to Gradual and Catastrophic Forgetting. In Findings of the Association for Computational Linguistics: ACL 2024, Lun-Wei Ku, Andre Martins, and Vivek Srikumar (Eds.). Association for Computational Linguistics, Bangkok, Thailand, 15202–15232. doi:10.18653/v1/2024.findings-acl.902 [77] Suchin Gururangan, Ana Marasović, Swabha Swayamdipta, Kyle Lo, Iz Beltagy, Doug Downey, and Noah A. Smith. 2020. Don’t Stop Pretraining: Adapt Language Models to Domains and Tasks. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, Dan Jurafsky, Joyce Chai, Natalie Schluter, and Joel Tetreault (Eds.). Association for Computational Linguistics, Online, 8342–8360. doi:10.18653/v1/2020.acl-main.740 [78] Bernal Jiménez Gutiérrez, Yiheng Shu, Yu Gu, Michihiro Yasunaga, and Yu Su. 2024. HippoRAG: neurobiologically inspired long-term memory for large language models. In Proceedings of the 38th International Conference on Neural Information Processing Systems (NIPS ’24, Vol. 37). Curran Associates Inc., Red Hook, NY, USA, 59532–59569. [79] Philipp Hacker and Matthias Holweg. 2026. The regulation of fine-tuning: Federated compliance for modified general-purpose AI models. Computer Law & Security Review 60 (April 2026), 106234. doi:10.1016/j.clsr.2025.106234 [80] Jialong Han, Si Zhang, and Ke Zhang. 2025. Dual Decomposition of Weights and Singular Value Low Rank Adaptation. https://arxiv.org/abs/2505.14367v2 [81] Song Han, Huizi Mao, and William J. Dally. 2016. Deep Compression: Compressing Deep Neural Networks with Pruning, Trained Quantization and Huffman Coding. doi:10.48550/arXiv.1510.00149 [82] Jason Hausenloy, Duncan McClements, and Madhavendra Thakur. 2025. Towards Data Governance of Frontier AI Models. doi:10.48550/arXiv.2412.03824 [83] Soufiane Hayou, Nikhil Ghosh, and Bin Yu. 2024. LoRA+: efficient low rank adaptation of large models. In Proceedings of the 41st International Conference on Machine Learning (ICML’24, Vol. 235). JMLR.org, Vienna, Austria, 17783–17806. [84] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. 2015. Distilling the Knowledge in a Neural Network. doi:10.48550/ arXiv.1503.02531 [85] Steven C. H. Hoi, Doyen Sahoo, Jing Lu, and Peilin Zhao. 2021. Online learning: A comprehensive survey. Neurocomputing 459 (Oct. 2021), 249–289. doi:10.1016/j.neucom.2021.04.112 [86] Jiwoo Hong, Noah Lee, and James Thorne. 2024. ORPO: Monolithic Preference Optimization without Reference Model. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, Yaser Al-Onaizan, Mohit Bansal, and Yun-Nung Chen (Eds.). Association for Computational Linguistics, Miami, Florida, USA, 11170–11189. doi:10.18653/v1/2024.emnlp-main.626 [87] Timothy Hospedales, Antreas Antoniou, Paul Micaelli, and Amos Storkey. 2022. Meta-Learning in Neural Networks: A Survey. IEEE Transactions on Pattern Analysis & Machine Intelligence 44, 09 (Sept. 2022), 5149–5169. doi:10.1109/ TPAMI.2021.3079209 [88] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. 2019. Parameter-Efficient Transfer Learning for NLP. In Proceedings of the 36th International Conference on Machine Learning (Proceedings of Machine Learning Research, Vol. 97), Kamalika Chaudhuri and Ruslan Salakhutdinov (Eds.). PMLR, 2790–2799. https://proceedings.mlr.press/v97/houlsby19a.html [89] Cheng-Hsun Hsueh, Paul Kuo-Ming Huang, Tzu-Han Lin, Che Wei Liao, Hung-Chieh Fang, Chao-Wei Huang, and Yun-Nung Chen. 2024. Editing the Mind of Giants: An In-Depth Exploration of Pitfalls of Knowledge Editing in Large Language Models. In Findings of the Association for Computational Linguistics: EMNLP 2024, Yaser Al-Onaizan, Mohit Bansal, and Yun-Nung Chen (Eds.). Association for Computational Linguistics, Miami, Florida, USA, 9417–9429. doi:10.18653/v1/2024.findings-emnlp.550 [90] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. LoRA: Low-Rank Adaptation of Large Language Models. https://www.microsoft.com/en-us/research/ publication/lora-low-rank-adaptation-of-large-language-models/ [91] Jikun Hu, Dongsheng Guo, Yuli Liu, Qingyao Ai, Lixuan Wang, Xuebing Sun, Qilei Zhang, Quan Zhou, and Cheng Luo. 2025. PEPE: Long-context Extension for Large Language Models via Periodic Extrapolation Positional Encodings. In Findings of the Association for Computational Linguistics: EMNLP 2025, Christos Christodoulopoulos, Tanmoy Chakraborty, Carolyn Rose, and Violet Peng (Eds.). Association for Computational Linguistics, Suzhou, China, 21075–21085. doi:10.18653/v1/2025.findings-emnlp.1149 [92] Jian Hu, Jason Klein Liu, Haotian Xu, and Wei Shen. 2025. REINFORCE++: Stabilizing Critic-Free Policy Optimization with Global Advantage Normalization. doi:10.48550/arXiv.2501.03262 [93] Shengyuan Hu, Yiwei Fu, Steven Wu, and Virginia Smith. 2024. Unlearning or Obfuscating? Jogging the Memory of Unlearned LLMs via Benign Relearning. https://openreview.net/forum?id=fMNRYBvcQN [94] Chengsong Huang, Qian Liu, Bill Yuchen Lin, Tianyu Pang, Chao Du, and Min Lin. 2024. LoraHub: Efficient Cross-Task Generalization via Dynamic LoRA Composition. https://openreview.net/forum?id=TrloAXEJ2B [95] Chenyu Huang, Peng Ye, Tao Chen, Tong He, Xiangyu Yue, and Wanli Ouyang. 2024. EMR-MERGING: tuning-free high-performance model merging. In Proceedings of the 38th International Conference on Neural Information Processing

34

Afdideh, Seoane, and Abtahi

Systems (NIPS ’24, Vol. 37). Curran Associates Inc., Red Hook, NY, USA, 122741–122769. [96] Yue Huang, Chujie Gao, Siyuan Wu, Haoran Wang, Xiangqi Wang, Yujun Zhou, Yanbo Wang, Jiayi Ye, Jiawen Shi, Qihui Zhang, Yuan Li, Han Bao, Zhaoyi Liu, Tianrui Guan, Dongping Chen, Ruoxi Chen, Kehan Guo, Andy Zou, Bryan Hooi Kuen-Yew, Caiming Xiong, Elias Stengel-Eskin, Hongyang Zhang, Hongzhi Yin, Huan Zhang, Huaxiu Yao, Jaehong Yoon, Jieyu Zhang, Kai Shu, Kaijie Zhu, Ranjay Krishna, Swabha Swayamdipta, Taiwei Shi, Weijia Shi, Xiang Li, Yiwei Li, Yuexing Hao, Zhihao Jia, Zhize Li, Xiuying Chen, Zhengzhong Tu, Xiyang Hu, Tianyi Zhou, Jieyu Zhao, Lichao Sun, Furong Huang, Or Cohen Sasson, Prasanna Sattigeri, Anka Reuel, Max Lamparth, Yue Zhao, Nouha Dziri, Yu Su, Huan Sun, Heng Ji, Chaowei Xiao, Mohit Bansal, Nitesh V. Chawla, Jian Pei, Jianfeng Gao, Michael Backes, Philip S. Yu, Neil Zhenqiang Gong, Pin-Yu Chen, Bo Li, Dawn Song, and Xiangliang Zhang. 2025. On the Trustworthiness of Generative Foundation Models: Guideline, Assessment, and Perspective. doi:10.48550/arXiv.2502.14296 [97] Dang Huu-Tien, Tin Pham, Hoang Thanh-Tung, and Naoya Inoue. 2025. On effects of steering latent representation for large language model unlearning. In Proceedings of the Thirty-Ninth AAAI Conference on Artificial Intelligence and Thirty-Seventh Conference on Innovative Applications of Artificial Intelligence and Fifteenth Symposium on Educational Advances in Artificial Intelligence (AAAI’25/IAAI’25/EAAI’25, Vol. 39). AAAI Press, 23733–23742. doi:10.1609/aaai. v39i22.34544 [98] Nam Hyeon-Woo, Moon Ye-Bin, and Tae-Hyun Oh. 2021. FedPara: Low-rank Hadamard Product for CommunicationEfficient Federated Learning. https://openreview.net/forum?id=d71n4ftoCBy [99] H. Daume Iii and D. Marcu. 2006. Domain Adaptation for Statistical Classifiers. Journal of Artificial Intelligence Research 26 (June 2006), 101–126. doi:10.1613/jair.1872 [100] Gabriel Ilharco, Marco Tulio Ribeiro, Mitchell Wortsman, Suchin Gururangan, Ludwig Schmidt, Hannaneh Hajishirzi, and Ali Farhadi. 2023. Editing Models with Task Arithmetic. doi:10.48550/arXiv.2212.04089 [101] Pavel Izmailov, Dmitrii Podoprikhin, Timur Garipov, Dmitry Vetrov, and Andrew Gordon Wilson. 2019. Averaging Weights Leads to Wider Optima and Better Generalization. doi:10.48550/arXiv.1803.05407 [102] Nicolas Jourdan, Tom Bayer, Tobias Biegel, and Joachim Metternich. 2023. Handling concept drift in deep learning applications for process monitoring. Procedia CIRP 120 (Jan. 2023), 33–38. doi:10.1016/j.procir.2023.08.007 [103] Peter Kairouz and H. Brendan McMahan. 2021. Advances and Open Problems in Federated Learning. Foundations and Trends in Machine Learning 14, 1-2 (June 2021), 1–210. doi:10.1561/2200000083 [104] Damjan Kalajdzievski. 2023. A Rank Stabilization Scaling Factor for Fine-Tuning with LoRA. https://arxiv.org/abs/ 2312.03732v1 [105] Vladimir Karpukhin, Barlas Oguz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. 2020. Dense Passage Retrieval for Open-Domain Question Answering. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), Bonnie Webber, Trevor Cohn, Yulan He, and Yang Liu (Eds.). Association for Computational Linguistics, Online, 6769–6781. doi:10.18653/v1/2020.emnlp-main.550 [106] Omar Khattab, Arnav Singhvi, Paridhi Maheshwari, Zhiyuan Zhang, Keshav Santhanam, Sri Vardhamanan, Saiful Haq, Ashutosh Sharma, Thomas T. Joshi, Hanna Moazam, Heather Miller, Matei Zaharia, and Christopher Potts. 2024. DSPy: Compiling Declarative Language Model Calls into Self-Improving Pipelines. doi:10.48550/arXiv.2310.03714 [107] Been Kim, Martin Wattenberg, Justin Gilmer, Carrie Cai, James Wexler, Fernanda Viegas, and Rory Sayres. 2018. Interpretability Beyond Feature Attribution: Quantitative Testing with Concept Activation Vectors (TCAV). In Proceedings of the 35th International Conference on Machine Learning. PMLR, 2668–2677. https://proceedings.mlr. press/v80/kim18d.html [108] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A. Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, Demis Hassabis, Claudia Clopath, Dharshan Kumaran, and Raia Hadsell. 2017. Overcoming catastrophic forgetting in neural networks. Proceedings of the National Academy of Sciences 114, 13 (March 2017), 3521–3526. doi:10.1073/pnas.1611835114 [109] Wouter Kool, Herke van Hoof, and Max Welling. 2019. Buy 4 REINFORCE Samples, Get a Baseline for Free! (April 2019). https://openreview.net/forum?id=r1lgTGL5DE [110] Dawid Jan Kopiczko, Tijmen Blankevoort, and Yuki M. Asano. 2023. VeRA: Vector-based Random Matrix Adaptation. https://openreview.net/forum?id=NjNfLdxr3A [111] Harrison Lee, Samrat Phatale, Hassan Mansoor, Thomas Mesnard, Johan Ferret, Kellie Ren Lu, Colton Bishop, Ethan Hall, Victor Carbune, Abhinav Rastogi, and Sushant Prakash. 2024. RLAIF vs. RLHF: Scaling Reinforcement Learning from Human Feedback with AI Feedback. In Proceedings of the 41st International Conference on Machine Learning (Proceedings of Machine Learning Research, Vol. 235), Ruslan Salakhutdinov, Zico Kolter, Katherine Heller, Adrian Weller, Nuria Oliver, Jonathan Scarlett, and Felix Berkenkamp (Eds.). PMLR, 26874–26901. [112] Karim Lekadir, Alejandro F. Frangi, Antonio R. Porras, Ben Glocker, Celia Cintas, Curtis P. Langlotz, Eva Weicken, Folkert W. Asselbergs, Fred Prior, Gary S. Collins, Georgios Kaissis, Gianna Tsakou, Irène Buvat, Jayashree KalpathyCramer, John Mongan, Julia A. Schnabel, Kaisar Kushibar, Katrine Riklund, Kostas Marias, Lameck M. Amugongo, Lauren A. Fromont, Lena Maier-Hein, Leonor Cerdá-Alberich, Luis Martí-Bonmatí, M. Jorge Cardoso, Maciej Bobowicz,

A Six-Dimensional Taxonomy of Post-Training Adaptation Techniques with Applications in AI Governance

35

Mahsa Shabani, Manolis Tsiknakis, Maria A. Zuluaga, Marie-Christine Fritzsche, Marina Camacho, Marius George Linguraru, Markus Wenzel, Marleen De Bruijne, Martin G. Tolsgaard, Melanie Goisauf, Mónica Cano Abadía, Nikolaos Papanikolaou, Noussair Lazrak, Oriol Pujol, Richard Osuala, Sandy Napel, Sara Colantonio, Smriti Joshi, Stefan Klein, Susanna Aussó, Wendy A. Rogers, Zohaib Salahuddin, Martijn P. A. Starmans, and FUTURE-AI Consortium. 2025. FUTURE-AI: international consensus guideline for trustworthy and deployable artificial intelligence in healthcare. BMJ 388 (Feb. 2025), e081554. doi:10.1136/bmj-2024-081554 [113] Brian Lester, Rami Al-Rfou, and Noah Constant. 2021. The Power of Scale for Parameter-Efficient Prompt Tuning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, Marie-Francine Moens, Xuanjing Huang, Lucia Specia, and Scott Wen-tau Yih (Eds.). Association for Computational Linguistics, Online and Punta Cana, Dominican Republic, 3045–3059. doi:10.18653/v1/2021.emnlp-main.243 [114] Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, Sebastian Riedel, and Douwe Kiela. 2020. Retrieval-augmented generation for knowledge-intensive NLP tasks. In Proceedings of the 34th International Conference on Neural Information Processing Systems (NIPS ’20). Curran Associates Inc., Red Hook, NY, USA, 9459–9474. [115] Dawei Li, Bohan Jiang, Liangjie Huang, Alimohammad Beigi, Chengshuai Zhao, Zhen Tan, Amrita Bhattacharjee, Yuxuan Jiang, Canyu Chen, Tianhao Wu, Kai Shu, Lu Cheng, and Huan Liu. 2025. From Generation to Judgment: Opportunities and Challenges of LLM-as-a-judge. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, Christos Christodoulopoulos, Tanmoy Chakraborty, Carolyn Rose, and Violet Peng (Eds.). Association for Computational Linguistics, Suzhou, China, 2757–2791. doi:10.18653/v1/2025.emnlp-main.138 [116] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. 2023. BLIP-2: Bootstrapping Language-Image Pre-training with Frozen Image Encoders and Large Language Models. In Proceedings of the 40th International Conference on Machine Learning (Proceedings of Machine Learning Research, Vol. 202), Andreas Krause, Emma Brunskill, Kyunghyun Cho, Barbara Engelhardt, Sivan Sabato, and Jonathan Scarlett (Eds.). PMLR, 19730–19742. https://proceedings.mlr. press/v202/li23q.html [117] Kenneth Li, Oam Patel, Fernanda Viégas, Hanspeter Pfister, and Martin Wattenberg. 2023. Inference-time intervention: eliciting truthful answers from a language model. In Proceedings of the 37th International Conference on Neural Information Processing Systems (NIPS ’23). Curran Associates Inc., Red Hook, NY, USA, 41451–41530. [118] Nathaniel Li, Alexander Pan, Anjali Gopal, Summer Yue, Daniel Berrios, Alice Gatti, Justin D. Li, Ann-Kathrin Dombrowski, Shashwat Goel, Gabriel Mukobi, Nathan Helm-Burger, Rassin Lababidi, Lennart Justen, Andrew Bo Liu, Michael Chen, Isabelle Barrass, Oliver Zhang, Xiaoyuan Zhu, Rishub Tamirisa, Bhrugu Bharathi, Ariel Herbert-Voss, Cort B. Breuer, Andy Zou, Mantas Mazeika, Zifan Wang, Palash Oswal, Weiran Lin, Adam Alfred Hunt, Justin Tienken-Harder, Kevin Y. Shih, Kemper Talley, John Guan, Ian Steneker, David Campbell, Brad Jokubaitis, Steven Basart, Stephen Fitz, Ponnurangam Kumaraguru, Kallol Krishna Karmakar, Uday Tupakula, Vijay Varadharajan, Yan Shoshitaishvili, Jimmy Ba, Kevin M. Esvelt, Alexandr Wang, and Dan Hendrycks. 2024. The WMDP Benchmark: Measuring and Reducing Malicious Use with Unlearning. In Proceedings of the 41st International Conference on Machine Learning. PMLR, 28525–28550. https://proceedings.mlr.press/v235/li24bc.html [119] Xiaopeng Li, Shasha Li, Shezheng Song, Jing Yang, Jun Ma, and Jie Yu. 2024. PMET: precise model editing in a transformer. In Proceedings of the Thirty-Eighth AAAI Conference on Artificial Intelligence and Thirty-Sixth Conference on Innovative Applications of Artificial Intelligence and Fourteenth Symposium on Educational Advances in Artificial Intelligence (AAAI’24/IAAI’24/EAAI’24, Vol. 38). AAAI Press, 18564–18572. doi:10.1609/aaai.v38i17.29818 [120] Xian Li, Ping Yu, Chunting Zhou, Timo Schick, Omer Levy, Luke Zettlemoyer, Jason E. Weston, and Mike Lewis. 2023. Self-Alignment with Instruction Backtranslation. https://openreview.net/forum?id=1oijHJBRsT [121] Xiang Lisa Li and Percy Liang. 2021. Prefix-Tuning: Optimizing Continuous Prompts for Generation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), Chengqing Zong, Fei Xia, Wenjie Li, and Roberto Navigli (Eds.). Association for Computational Linguistics, Online, 4582–4597. doi:10.18653/v1/2021.acl-long.353 [122] Zhenguo Li, Fengwei Zhou, Fei Chen, and Hang Li. 2017. Meta-SGD: Learning to Learn Quickly for Few-Shot Learning. doi:10.48550/arXiv.1707.09835 [123] Hunter Lightman, Vineet Kosaraju, Yuri Burda, Harrison Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. 2024. Let’s Verify Step by Step. In International Conference on Learning Representations (ICLR). https://openreview.net/forum?id=v8L0pN6EOi [124] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. 2023. Visual Instruction Tuning. In Advances in Neural Information Processing Systems, A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine (Eds.), Vol. 36. Curran Associates, Inc., 34892–34916. https://proceedings.neurips.cc/paper_files/paper/2023/file/ 6dcf277ea32ce3288914faf369fe6de0-Paper-Conference.pdf [125] Haokun Liu, Derek Tam, Mohammed Muqeeth, Jay Mohta, Tenghao Huang, Mohit Bansal, and Colin Raffel. 2022. Few-shot parameter-efficient fine-tuning is better and cheaper than in-context learning. In Proceedings of the 36th

36

Afdideh, Seoane, and Abtahi

International Conference on Neural Information Processing Systems (NIPS ’22). Curran Associates Inc., Red Hook, NY, USA, 1950–1965. [126] Pengfei Liu, Weizhe Yuan, Jinlan Fu, Zhengbao Jiang, Hiroaki Hayashi, and Graham Neubig. 2023. Pre-train, Prompt, and Predict: A Systematic Survey of Prompting Methods in Natural Language Processing. ACM Comput. Surv. 55, 9 (Jan. 2023), 195:1–195:35. doi:10.1145/3560815 [127] Shih-Yang Liu, Chien-Yi Wang, Hongxu Yin, Pavlo Molchanov, Yu-Chiang Frank Wang, Kwang-Ting Cheng, and Min-Hung Chen. 2024. DoRA: weight-decomposed low-rank adaptation. In Proceedings of the 41st International Conference on Machine Learning (ICML’24, Vol. 235). JMLR.org, Vienna, Austria, 32100–32121. [128] Lin Long, Rui Wang, Ruixuan Xiao, Junbo Zhao, Xiao Ding, Gang Chen, and Haobo Wang. 2024. On LLMs-Driven Synthetic Data Generation, Curation, and Evaluation: A Survey. In Findings of the Association for Computational Linguistics: ACL 2024, Lun-Wei Ku, Andre Martins, and Vivek Srikumar (Eds.). Association for Computational Linguistics, Bangkok, Thailand, 11065–11082. doi:10.18653/v1/2024.findings-acl.658 [129] Jie Lu, Anjin Liu, Fan Dong, Feng Gu, João Gama, and Guangquan Zhang. 2019. Learning under Concept Drift: A Review. IEEE Transactions on Knowledge and Data Engineering 31, 12 (Dec. 2019), 2346–2363. doi:10.1109/TKDE.2018. 2876857 [130] Pengqian Lu, Jie Lu, Anjin Liu, and Guangquan Zhang. 2025. Early concept drift detection via prediction uncertainty. In Proceedings of the Thirty-Ninth AAAI Conference on Artificial Intelligence and Thirty-Seventh Conference on Innovative Applications of Artificial Intelligence and Fifteenth Symposium on Educational Advances in Artificial Intelligence (AAAI’25/IAAI’25/EAAI’25, Vol. 39). AAAI Press, 19124–19132. doi:10.1609/aaai.v39i18.34105 [131] Wei Lu, Rachel K. Luu, and Markus J. Buehler. 2025. Fine-tuning large language models for domain adaptation: exploration of training strategies, scaling, model merging and synergistic capabilities. npj Comput Mater 11, 1 (March 2025), 84. doi:10.1038/s41524-025-01564-y [132] Tongxu Luo, Jiahe Lei, Fangyu Lei, Weihao Liu, Shizhu He, Jun Zhao, and Kang Liu. 2024. MoELoRA: Contrastive Learning Guided Mixture of Experts on Parameter-Efficient Fine-Tuning for Large Language Models. doi:10.48550/ arXiv.2402.12851 [133] Shumin Ma, Zhiri Yuan, Qi Wu, Yiyan Huang, Xixu Hu, Cheuk Hang Leung, Dongdong Wang, and Zhixiang Huang. 2024. Deep Into the Domain Shift: Transfer Learning Through Dependence Regularization. IEEE Transactions on Neural Networks and Learning Systems 35, 10 (Oct. 2024), 14409–14423. doi:10.1109/TNNLS.2023.3279099 [134] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. 2018. Towards Deep Learning Models Resistant to Adversarial Attacks. https://openreview.net/forum?id=rJzIBfZAb [135] Michael Matena and Colin Raffel. 2022. Merging models with fisher-weighted averaging. In Proceedings of the 36th International Conference on Neural Information Processing Systems (NIPS ’22). Curran Associates Inc., Red Hook, NY, USA, 17703–17716. [136] Michael McCloskey and Neal J. Cohen. 1989. Catastrophic Interference in Connectionist Networks: The Sequential Learning Problem. Psychology of Learning and Motivation, Vol. 24. Academic Press, 109–165. doi:10.1016/S00797421(08)60536-8 [137] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. 2017. CommunicationEfficient Learning of Deep Networks from Decentralized Data. In Proceedings of the 20th International Conference on Artificial Intelligence and Statistics (Proceedings of Machine Learning Research, Vol. 54), Aarti Singh and Jerry Zhu (Eds.). PMLR, 1273–1282. https://proceedings.mlr.press/v54/mcmahan17a.html [138] Viraj Mehta, Abhinav Komanduri, Rishabh Singh Bhadouriya, Vilina Mehta, Michael David Johnson, Priyanka Shrestha, Margaret Nikolov, Bhav Jain, Nigam Shah, and Kevin Schulman. 2025. Evaluating transparency in AI/ML model characteristics for FDA-reviewed medical devices. npj Digit. Med. 8, 1 (Nov. 2025), 673. doi:10.1038/s41746025-02052-9 [139] Fanxu Meng, Zhaohui Wang, and Muhan Zhang. 2024. PiSSA: Principal Singular Values and Singular Vectors Adaptation of Large Language Models. https://openreview.net/forum?id=6ZBHIEtdP4 [140] Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov. 2022. Locating and Editing Factual Associations in GPT. In Advances in Neural Information Processing Systems, S. Koyejo, S. Mohamed, A. Agarwal, D. Belgrave, K. Cho, and A. Oh (Eds.), Vol. 35. Curran Associates, Inc., 17359–17372. https://proceedings.neurips.cc/paper_files/paper/ 2022/file/6f1d43d5a82a37e89b0665b33bf3a182-Paper-Conference.pdf [141] Kevin Meng, Arnab Sen Sharma, Alex Andonian, Yonatan Belinkov, and David Bau. 2023. Mass-Editing Memory in a Transformer. [142] Yu Meng, Mengzhou Xia, and Danqi Chen. 2024. SimPO: Simple Preference Optimization with a ReferenceFree Reward. https://openreview.net/forum?id=3Tzcot1LKb&referrer=%5Bthe%20profile%20of%20Danqi%20Chen% 5D(%2Fprofile%3Fid%3D~Danqi_Chen1) [143] Tomas Mikolov, Wen-tau Yih, and Geoffrey Zweig. 2013. Linguistic Regularities in Continuous Space Word Representations. In Proceedings of the 2013 Conference of the North American Chapter of the Association for Computational

A Six-Dimensional Taxonomy of Post-Training Adaptation Techniques with Applications in AI Governance

37

Linguistics: Human Language Technologies, Lucy Vanderwende, Hal Daumé III, and Katrin Kirchhoff (Eds.). Association for Computational Linguistics, Atlanta, Georgia, 746–751. https://aclanthology.org/N13-1090/ [144] MiniMax, Aili Chen, Aonian Li, Bangwei Gong, Binyang Jiang, Bo Fei, Bo Yang, Boji Shan, Changqing Yu, Chao Wang, Cheng Zhu, Chengjun Xiao, Chengyu Du, Chi Zhang, Chu Qiao, Chunhao Zhang, Chunhui Du, Congchao Guo, Da Chen, Deming Ding, Dianjun Sun, Dong Li, Enwei Jiao, Haigang Zhou, Haimo Zhang, Han Ding, Haohai Sun, Haoyu Feng, Huaiguang Cai, Haichao Zhu, Jian Sun, Jiaqi Zhuang, Jiaren Cai, Jiayuan Song, Jin Zhu, Jingyang Li, Jinhao Tian, Jinli Liu, Junhao Xu, Junjie Yan, Junteng Liu, Junxian He, Kaiyi Feng, Ke Yang, Kecheng Xiao, Le Han, Leyang Wang, Lianfei Yu, Liheng Feng, Lin Li, Lin Zheng, Linge Du, Lingyu Yang, Lunbin Zeng, Minghui Yu, Mingliang Tao, Mingyuan Chi, Mozhi Zhang, Mujie Lin, Nan Hu, Nongyu Di, Peng Gao, Pengfei Li, Pengyu Zhao, Qibing Ren, Qidi Xu, Qile Li, Qin Wang, Rong Tian, Ruitao Leng, Shaoxiang Chen, Shaoyu Chen, Shengmin Shi, Shitong Weng, Shuchang Guan, Shuqi Yu, Sichen Li, Songquan Zhu, Tengfei Li, Tianchi Cai, Tianrun Liang, Weiyu Cheng, Weize Kong, Wenkai Li, Xiancai Chen, Xiangjun Song, Xiao Luo, Xiao Su, Xiaobo Li, Xiaodong Han, Xinzhu Hou, Xuan Lu, Xun Zou, Xuyang Shen, Yan Gong, Yan Ma, Yang Wang, Yiqi Shi, Yiran Zhong, Yonghong Duan, Yongxiang Fu, Yongyi Hu, Yu Gao, Yuanxiang Fan, Yufeng Yang, Yuhao Li, Yulin Hu, Yunan Huang, Yunji Li, Yunzhi Xu, Yuxin Mao, Yuxuan Shi, Yuze Wenren, Zehan Li, Zelin Li, Zhanxu Tian, Zhengmao Zhu, Zhenhua Fan, Zhenzhen Wu, Zhichao Xu, Zhihang Yu, Zhiheng Lyu, Zhuo Jiang, Zibo Gao, Zijia Wu, Zijian Song, and Zijun Sun. 2025. MiniMax-M1: Scaling Test-Time Compute Efficiently with Lightning Attention. doi:10.48550/arXiv.2506.13585 [145] Eric Mitchell, Charles Lin, Antoine Bosselut, Chelsea Finn, and Christopher D. Manning. 2022. Fast Model Editing at Scale. In International Conference on Learning Representations (ICLR). https://openreview.net/pdf?id=0DcZxeWfOPt [146] Eric Mitchell, Charles Lin, Antoine Bosselut, Christopher D Manning, and Chelsea Finn. 2022. Memory-Based Model Editing at Scale. In Proceedings of the 39th International Conference on Machine Learning (Proceedings of Machine Learning Research, Vol. 162), Kamalika Chaudhuri, Stefanie Jegelka, Le Song, Csaba Szepesvari, Gang Niu, and Sivan Sabato (Eds.). PMLR, 15817–15831. https://proceedings.mlr.press/v162/mitchell22a.html [147] Tom M. Mitchell. 1997. Machine Learning. McGraw-Hill. [148] Jose G. Moreno-Torres, Troy Raeder, Rocío Alaiz-Rodríguez, Nitesh V. Chawla, and Francisco Herrera. 2012. A unifying view on dataset shift in classification. Pattern Recognition 45, 1 (Jan. 2012), 521–530. doi:10.1016/j.patcog.2011.06.019 [149] Urs J. Muehlematter, Paola Daniore, and Kerstin N. Vokinger. 2021. Approval of artificial intelligence and machine learning-based medical devices in the USA and Europe (2015–20): a comparative analysis. The Lancet Digital Health 3, 3 (March 2021), e195–e203. doi:10.1016/S2589-7500(20)30292-2 [150] Jessica Newman. 2023. A Taxonomy of Trustworthiness for Artificial Intelligence. Technical Report. UC Berkeley Center for Long-Term Cybersecurity (CLTC). https://cltc.berkeley.edu/publication/a-taxonomy-of-trustworthinessfor-artificial-intelligence/ [151] Alex Nichol, Joshua Achiam, and John Schulman. 2018. On First-Order Meta-Learning Algorithms. doi:10.48550/ arXiv.1803.02999 [152] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul F Christiano, Jan Leike, and Ryan Lowe. 2022. Training language models to follow instructions with human feedback. In Advances in Neural Information Processing Systems, S. Koyejo, S. Mohamed, A. Agarwal, D. Belgrave, K. Cho, and A. Oh (Eds.), Vol. 35. Curran Associates, Inc., 27730–27744. https: //proceedings.neurips.cc/paper_files/paper/2022/file/b1efde53be364a73914f58805a001731-Paper-Conference.pdf [153] Sinno Jialin Pan and Qiang Yang. 2010. A Survey on Transfer Learning. IEEE Transactions on Knowledge and Data Engineering 22, 10 (Oct. 2010), 1345–1359. doi:10.1109/TKDE.2009.191 [154] Joon Sung Park, Joseph O’Brien, Carrie Jun Cai, Meredith Ringel Morris, Percy Liang, and Michael S. Bernstein. 2023. Generative Agents: Interactive Simulacra of Human Behavior. In Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology (UIST ’23). Association for Computing Machinery, New York, NY, USA, 1–22. doi:10.1145/3586183.3606763 [155] Seongsik Park, Sangmin Park, Jaieun Kim, and Harksoo Kim. 2025. MAKE: Memory-Associated Knowledge Editing. Transactions of the Association for Computational Linguistics 13 (2025), 938–952. doi:10.1162/tacl.a.26 [156] Bowen Peng, Jeffrey Quesnelle, Honglu Fan, and Enrico Shippole. 2023. YaRN: Efficient Context Window Extension of Large Language Models. https://openreview.net/forum?id=wHBfxhZu1u [157] Jonas Pfeiffer, Aishwarya Kamath, Andreas Rücklé, Kyunghyun Cho, and Iryna Gurevych. 2021. AdapterFusion: Non-Destructive Task Composition for Transfer Learning. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, Paola Merlo, Jorg Tiedemann, and Reut Tsarfaty (Eds.). Association for Computational Linguistics, Online, 487–503. doi:10.18653/v1/2021.eacl-main.39 [158] J. Platt. 1999. Probabilistic Outputs for Support vector Machines and Comparisons to Regularized Likelihood Methods. https://www.semanticscholar.org/paper/Probabilistic-Outputs-for-Support-vector-MachinesPlatt/42e5ed832d4310ce4378c44d05570439df28a393

38

Afdideh, Seoane, and Abtahi

[159] Lorena Poenaru-Olaru, June Sallou, Luis Cruz, Jan S. Rellermeyer, and Arie van Deursen. 2023. Retrain AI Systems Responsibly! Use Sustainable Concept Drift Adaptation Techniques. In 2023 IEEE/ACM 7th International Workshop on Green And Sustainable Software (GREENS). 17–18. doi:10.1109/GREENS59328.2023.00009 [160] Joris Postmus and Steven Abreu. 2024. Steering Large Language Models using Conceptors: Improving Addition-Based Activation Engineering. https://openreview.net/forum?id=gyAnAq16HC [161] Prompt Engineering Guide. 2026. Context Engineering Guide – Nextra. https://www.promptingguide.ai/guides/ context-engineering-guide [162] Stephan Rabanser, Stephan Günnemann, and Zachary Lipton. 2019. Failing Loudly: An Empirical Study of Methods for Detecting Dataset Shift. In Advances in Neural Information Processing Systems, H. Wallach, H. Larochelle, A. Beygelzimer, F. d’ Alché-Buc, E. Fox, and R. Garnett (Eds.), Vol. 32. Curran Associates, Inc. https://proceedings. neurips.cc/paper_files/paper/2019/file/846c260d715e5b854ffad5f70a516c88-Paper.pdf [163] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. 2021. Learning Transferable Visual Models From Natural Language Supervision. In Proceedings of the 38th International Conference on Machine Learning (Proceedings of Machine Learning Research, Vol. 139), Marina Meila and Tong Zhang (Eds.). PMLR, 8748–8763. https: //proceedings.mlr.press/v139/radford21a.html [164] Rafael Rafailov, Archit Sharma, Eric Mitchell, Stefano Ermon, Christopher D. Manning, and Chelsea Finn. 2023. Direct preference optimization: your language model is secretly a reward model. In Proceedings of the 37th International Conference on Neural Information Processing Systems (NIPS ’23). Curran Associates Inc., Red Hook, NY, USA, 53728– 53741. [165] Rajat Raina, Alexis Battle, Honglak Lee, Benjamin Packer, and Andrew Y. Ng. 2007. Self-taught learning: transfer learning from unlabeled data. In Proceedings of the 24th international conference on Machine learning (ICML ’07). Association for Computing Machinery, New York, NY, USA, 759–766. doi:10.1145/1273496.1273592 [166] Nina Rimsky, Nick Gabrieli, Julian Schulz, Meg Tong, Evan Hubinger, and Alexander Turner. 2024. Steering Llama 2 via Contrastive Activation Addition. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), Lun-Wei Ku, Andre Martins, and Vivek Srikumar (Eds.). Association for Computational Linguistics, Bangkok, Thailand, 15504–15522. doi:10.18653/v1/2024.acl-long.828 [167] Stephen Robertson and Hugo Zaragoza. 2009. The Probabilistic Relevance Framework: BM25 and Beyond. Foundations and Trends in Information Retrieval 4, 1-2 (Sept. 2009), 1–174. doi:10.1561/1500000019 [168] Andrei A. Rusu, Dushyant Rao, Jakub Sygnowski, Oriol Vinyals, Razvan Pascanu, Simon Osindero, and Raia Hadsell. 2018. Meta-Learning with Latent Embedding Optimization. https://openreview.net/forum?id=BJgklhAcK7 [169] Philipp Schmid. 2025. The New Skill in AI is Not Prompting, It’s Context Engineering. https://www.philschmid.de/ context-engineering [170] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. 2017. Proximal Policy Optimization Algorithms. doi:10.48550/arXiv.1707.06347 [171] D. Sculley, Gary Holt, Daniel Golovin, Eugene Davydov, Todd Phillips, Dietmar Ebner, Vinay Chaudhary, Michael Young, Jean-François Crespo, and Dan Dennison. 2015. Hidden Technical Debt in Machine Learning Systems. In Advances in Neural Information Processing Systems, Vol. 28. Curran Associates, Inc. https://papers.nips.cc/paper_ files/paper/2015/hash/86df7dcfd896fcaf2674f757a2463eba-Abstract.html [172] Parker Seegmiller and Sarah Masud Preum. 2026. Measuring Distribution Shift in User Prompts and Its Effects on LLM Performance. doi:10.48550/arXiv.2604.17650 [173] Rico Sennrich, Barry Haddow, and Alexandra Birch. 2016. Improving Neural Machine Translation Models with Monolingual Data. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), Katrin Erk and Noah A. Smith (Eds.). Association for Computational Linguistics, Berlin, Germany, 86–96. doi:10.18653/v1/P16-1009 [174] Burr Settles. 2009. Active Learning Literature Survey. Technical Report. University of Wisconsin-Madison Department of Computer Sciences. https://minds.wisconsin.edu/handle/1793/60660 [175] Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, Y. K. Li, Y. Wu, and Daya Guo. 2024. DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models. (April 2024). doi:10.48550/arXiv.2402.03300 [176] Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. 2017. Outrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts Layer. In International Conference on Learning Representations (ICLR). [177] Ying Sheng, Shiyi Cao, Dacheng Li, Coleman Hooper, Nicholas Lee, Shuo Yang, Christopher Chou, Banghua Zhu, Lianmin Zheng, Kurt Keutzer, Joseph E. Gonzalez, and Ion Stoica. 2024. S-LoRA: Serving Thousands of Concurrent LoRA Adapters. https://proceedings.mlsys.org/paper_files/paper/2024/file/906419cd502575b617cc489a1a696a67Paper-Conference.pdf

A Six-Dimensional Taxonomy of Post-Training Adaptation Techniques with Applications in AI Governance

39

[178] Haizhou Shi, Zihao Xu, Hengyi Wang, Weiyi Qin, Wenyuan Wang, Yibin Wang, Zifeng Wang, Sayna Ebrahimi, and Hao Wang. 2025. Continual Learning of Large Language Models: A Comprehensive Survey. ACM Comput. Surv. 58, 5 (Nov. 2025), 120:1–120:42. doi:10.1145/3735633 [179] Hidetoshi Shimodaira. 2000. Improving predictive inference under covariate shift by weighting the log-likelihood function. Journal of Statistical Planning and Inference 90, 2 (Oct. 2000), 227–244. doi:10.1016/S0378-3758(00)00115-4 [180] Ken Shoemake. 1985. Animating rotation with quaternion curves. SIGGRAPH Comput. Graph. 19, 3 (July 1985), 245–254. doi:10.1145/325165.325242 [181] Connor Shorten and Taghi M. Khoshgoftaar. 2019. A survey on Image Data Augmentation for Deep Learning. J Big Data 6, 1 (July 2019), 60. doi:10.1186/s40537-019-0197-0 [182] David Silver, Julian Schrittwieser, Karen Simonyan, Ioannis Antonoglou, Aja Huang, Arthur Guez, Thomas Hubert, Lucas Baker, Matthew Lai, Adrian Bolton, Yutian Chen, Timothy Lillicrap, Fan Hui, Laurent Sifre, George van den Driessche, Thore Graepel, and Demis Hassabis. 2017. Mastering the game of Go without human knowledge. Nature 550, 7676 (Oct. 2017), 354–359. doi:10.1038/nature24270 [183] Aditi Singh, Abul Ehtesham, Saket Kumar, Tala Talaei Khoei, and Athanasios V. Vasilakos. 2026. Agentic RetrievalAugmented Generation: A Survey on Agentic RAG. (April 2026). doi:10.48550/arXiv.2501.09136 [184] Charlie Victor Snell, Jaehoon Lee, Kelvin Xu, and Aviral Kumar. 2024. Scaling LLM Test-Time Compute Optimally Can be More Effective than Scaling Parameters for Reasoning. https://openreview.net/forum?id=4FWAwZtd2n [185] Jake Snell, Kevin Swersky, and Richard Zemel. 2017. Prototypical networks for few-shot learning. In Proceedings of the 31st International Conference on Neural Information Processing Systems (NIPS’17). Curran Associates Inc., Red Hook, NY, USA, 4080–4090. https://dl.acm.org/doi/10.5555/3294996.3295163 [186] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. 2020. FixMatch: Simplifying Semi-Supervised Learning with Consistency and Confidence. In Advances in Neural Information Processing Systems, H. Larochelle, M. Ranzato, R. Hadsell, M. F. Balcan, and H. Lin (Eds.), Vol. 33. Curran Associates, Inc., 596–608. https://proceedings.neurips.cc/paper_files/paper/2020/ file/06964dce9addb1c5cb5d6e3d9838f733-Paper.pdf [187] Martin Stanley and Raymond Sheh. 2026. Concept Note: AI RMF Profile on Trustworthy AI in Critical Infrastructure. National Institute of Standards and Technology (NIST) (April 2026). https://www.nist.gov/programs-projects/conceptnote-ai-rmf-profile-trustworthy-ai-critical-infrastructure [188] Monika Steidl, Michael Felderer, and Rudolf Ramler. 2023. The pipeline for the continuous development of artificial intelligence models—Current state of research and practice. Journal of Systems and Software 199 (May 2023), 111615. doi:10.1016/j.jss.2023.111615 [189] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. 2024. RoFormer: Enhanced transformer with Rotary Position Embedding. Neurocomputing 568 (Feb. 2024), 127063. doi:10.1016/j.neucom.2023.127063 [190] Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. 2020. Test-Time Training with Self-Supervision for Generalization under Distribution Shifts. In Proceedings of the 37th International Conference on Machine Learning (Proceedings of Machine Learning Research, Vol. 119), Hal Daumé III and Aarti Singh (Eds.). PMLR, 9229–9248. https://proceedings.mlr.press/v119/sun20b.html [191] Yi-Lin Sung, Jaemin Cho, and Mohit Bansal. 2022. VL-ADAPTER: Parameter-Efficient Transfer Learning for Visionand-Language Tasks. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 5217–5227. ISSN: 2575-7075. doi:10.1109/CVPR52688.2022.00516 [192] Matteo Testi, Matteo Ballabio, Emanuele Frontoni, Giulio Iannello, Sara Moccia, Paolo Soda, and Gennaro Vessio. 2022. MLOps: A Taxonomy and a Methodology. IEEE Access 10 (2022), 63606–63618. doi:10.1109/ACCESS.2022.3181730 [193] Sebastian Thrun and Lorien Pratt (Eds.). 1998. Learning to Learn. Springer US, Boston, MA. doi:10.1007/978-1-46155529-2 [194] Eric Todd, Millicent Li, Arnab Sen Sharma, Aaron Mueller, Byron C. Wallace, and David Bau. 2023. Function Vectors in Large Language Models. https://openreview.net/forum?id=AwyxtyMwaG [195] Dimitris Tsipras, Shibani Santurkar, Logan Engstrom, Alexander Turner, and Aleksander Madry. 2018. Robustness May Be at Odds with Accuracy. https://openreview.net/forum?id=SyxAb30cY7 [196] Alexander Matt Turner, Lisa Thiergart, Gavin Leech, David Udell, Juan J. Vazquez, Ulisse Mini, and Monte MacDiarmid. 2024. Steering Language Models With Activation Engineering. (Oct. 2024). doi:10.48550/arXiv.2308.10248 [197] U.S. Food and Drug Administration. 2025. Marketing Submission Recommendations for a Predetermined Change Control Plan for Artificial Intelligence-Enabled Device Software Functions. (Aug. 2025). https://www.fda.gov/regulatory-information/search-fda-guidance-documents/marketing-submissionrecommendations-predetermined-change-control-plan-artificial-intelligence [198] Mojtaba Valipour, Mehdi Rezagholizadeh, Ivan Kobyzev, and Ali Ghodsi. 2023. DyLoRA: Parameter-Efficient Tuning of Pre-trained Models using Dynamic Search-Free Low-Rank Adaptation. In Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, Andreas Vlachos and Isabelle Augenstein (Eds.).

40

Afdideh, Seoane, and Abtahi

Association for Computational Linguistics, Dubrovnik, Croatia, 3274–3287. doi:10.18653/v1/2023.eacl-main.239 [199] Gido M. van de Ven, Tinne Tuytelaars, and Andreas S. Tolias. 2022. Three types of incremental learning. Nat Mach Intell 4, 12 (Dec. 2022), 1185–1197. doi:10.1038/s42256-022-00568-3 [200] Apostol Vassilev, Alina Oprea, Alie Fordyce, Hyrum Anderson, Xander Davies, and Maia Hamin. 2025. Adversarial Machine Learning: A Taxonomy and Terminology of Attacks and Mitigations. National Institute of Standards and Technology (March 2025). doi:10.6028/NIST.AI.100-2e2025 [201] Oriol Vinyals, Charles Blundell, Timothy Lillicrap, Koray Kavukcuoglu, and Daan Wierstra. 2016. Matching networks for one shot learning. In Proceedings of the 30th International Conference on Neural Information Processing Systems (NIPS’16). Curran Associates Inc., Red Hook, NY, USA, 3637–3645. https://dl.acm.org/doi/10.5555/3157382.3157504 [202] Dequan Wang, Evan Shelhamer, Shaoteng Liu, B. Olshausen, and Trevor Darrell. 2021. Tent: Fully Test-Time Adaptation by Entropy Minimization. https://www.semanticscholar.org/paper/Tent%3A-Fully-Test-Time-Adaptation-byEntropy-Wang-Shelhamer/180c78b132f6369a384d22a9529551d86c8788d3 [203] Hanqing Wang, Yixia Li, Shuo Wang, Guanhua Chen, and Yun Chen. 2025. MiLoRA: Harnessing Minor Singular Components for Parameter-Efficient LLM Finetuning. In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), Luis Chiruzzo, Alan Ritter, and Lu Wang (Eds.). Association for Computational Linguistics, Albuquerque, New Mexico, 4823–4836. doi:10.18653/v1/2025.naacl-long.248 [204] Jindong Wang, Cuiling Lan, Chang Liu, Yidong Ouyang, Tao Qin, Wang Lu, Yiqiang Chen, Wenjun Zeng, and Philip S. Yu. 2023. Generalizing to Unseen Domains: A Survey on Domain Generalization. IEEE Transactions on Knowledge and Data Engineering 35, 8 (Aug. 2023), 8052–8072. doi:10.1109/TKDE.2022.3178128 [205] Ke Wang, Jiahui Zhu, Minjie Ren, Zeming Liu, Shiwei Li, Zongye Zhang, Chenkai Zhang, Xiaoyu Wu, Qiqi Zhan, Qingjie Liu, and Yunhong Wang. 2024. A Survey on Data Synthesis and Augmentation for Large Language Models. doi:10.48550/arXiv.2410.12896 [206] Liyuan Wang, Xingxing Zhang, Hang Su, and Jun Zhu. 2024. A Comprehensive Survey of Continual Learning: Theory, Method and Application. IEEE Transactions on Pattern Analysis and Machine Intelligence 46, 8 (Aug. 2024), 5362–5383. doi:10.1109/TPAMI.2024.3367329 [207] Peng Wang, Zexi Li, Ningyu Zhang, Ziwen Xu, Yunzhi Yao, Yong Jiang, Pengjun Xie, Fei Huang, and Huajun Chen. 2024. WISE: rethinking the knowledge memory for lifelong model editing of large language models. In Proceedings of the 38th International Conference on Neural Information Processing Systems (NIPS ’24, Vol. 37). Curran Associates Inc., Red Hook, NY, USA, 53764–53797. [208] Yue Wang, Dehang Fu, Jie Tan, Junxiao Han, Yao Wan, Lixin Cui, Lu Bai, and Philip S. Yu. 2025. Detecting Intent Drift in Continuous Conversation via Temporal Transition Accumulation. In 2025 IEEE International Conference on Data Mining (ICDM). 773–782. doi:10.1109/ICDM65498.2025.00085 [209] Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A. Smith, Daniel Khashabi, and Hannaneh Hajishirzi. 2023. Self-Instruct: Aligning Language Models with Self-Generated Instructions. doi:10.48550/arXiv.2212.10560 [210] Zhengbo Wang, Jian Liang, Ran He, Zilei Wang, and Tieniu Tan. 2024. LoRA-Pro: Are Low-Rank Adapters Properly Optimized? https://openreview.net/forum?id=gTwRMU3lJ5 [211] Jan Wehner, Sahar Abdelnabi, Daniel Tan, David Krueger, and Mario Fritz. 2025. Taxonomy, Opportunities, and Challenges of Representation Engineering for Large Language Models. Transactions on Machine Learning Research (April 2025). https://openreview.net/forum?id=2U1KIfmaU9 [212] Jason Wei and Kai Zou. 2019. EDA: Easy Data Augmentation Techniques for Boosting Performance on Text Classification Tasks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), Kentaro Inui, Jing Jiang, Vincent Ng, and Xiaojun Wan (Eds.). Association for Computational Linguistics, Hong Kong, China, 6382–6388. doi:10.18653/v1/D19-1670 [213] Gerhard Widmer and Miroslav Kubat. 1996. Learning in the presence of concept drift and hidden contexts. Mach Learn 23, 1 (April 1996), 69–101. doi:10.1007/BF00116900 [214] Mitchell Wortsman, Gabriel Ilharco, Samir Ya Gadre, Rebecca Roelofs, Raphael Gontijo-Lopes, Ari S Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, and Ludwig Schmidt. 2022. Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. In Proceedings of the 39th International Conference on Machine Learning (Proceedings of Machine Learning Research, Vol. 162), Kamalika Chaudhuri, Stefanie Jegelka, Le Song, Csaba Szepesvari, Gang Niu, and Sivan Sabato (Eds.). PMLR, 23965–23998. https:// proceedings.mlr.press/v162/wortsman22a.html [215] Eric Wu, Kevin Wu, Roxana Daneshjou, David Ouyang, Daniel E. Ho, and James Zou. 2021. How medical AI devices are evaluated: limitations and recommendations from an analysis of FDA approvals. Nat Med 27, 4 (April 2021), 582–584. doi:10.1038/s41591-021-01312-x

A Six-Dimensional Taxonomy of Post-Training Adaptation Techniques with Applications in AI Governance

41

[216] Zhengxuan Wu, Aryaman Arora, Zheng Wang, Atticus Geiger, Dan Jurafsky, Christopher D. Manning, and Christopher Potts. 2024. ReFT: Representation Finetuning for Language Models. In Advances in Neural Information Processing Systems, Vol. 37. Curran Associates, Inc., 63908–63962. doi:10.52202/079017-2041 [217] Haoran Xu, Amr Sharaf, Yunmo Chen, Weiting Tan, Lingfeng Shen, Benjamin Van Durme, Kenton Murray, and Young Jin Kim. 2024. Contrastive preference optimization: pushing the boundaries of LLM performance in machine translation. In Proceedings of the 41st International Conference on Machine Learning (ICML’24, Vol. 235). JMLR.org, Vienna, Austria, 55204–55224. [218] Prateek Yadav, Derek Tam, Leshem Choshen, Colin A Raffel, and Mohit Bansal. 2023. TIES-Merging: Resolving Interference When Merging Models. In Advances in Neural Information Processing Systems, A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine (Eds.), Vol. 36. Curran Associates, Inc., 7093–7115. https:// proceedings.neurips.cc/paper_files/paper/2023/file/1644c9af28ab7916874f6fd6228a9bcf-Paper-Conference.pdf [219] Chengrun Yang, Xuezhi Wang, Yifeng Lu, Hanxiao Liu, Quoc V. Le, Denny Zhou, and Xinyun Chen. 2024. Large Language Models as Optimizers. doi:10.48550/arXiv.2309.03409 [220] Jin Yao, Eli Chien, Minxin Du, Xinyao Niu, Tianhao Wang, Zezhou Cheng, and Xiang Yue. 2024. Machine Unlearning of Pre-trained Large Language Models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), Lun-Wei Ku, Andre Martins, and Vivek Srikumar (Eds.). Association for Computational Linguistics, Bangkok, Thailand, 8403–8419. doi:10.18653/v1/2024.acl-long.457 [221] Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Thomas L. Griffiths, Yuan Cao, and Karthik Narasimhan. 2023. Tree of thoughts: deliberate problem solving with large language models. In Proceedings of the 37th International Conference on Neural Information Processing Systems (NIPS ’23). Curran Associates Inc., Red Hook, NY, USA, 11809–11822. [222] Jason Yosinski, Jeff Clune, Yoshua Bengio, and Hod Lipson. 2014. How transferable are features in deep neural networks?. In Advances in Neural Information Processing Systems, Z. Ghahramani, M. Welling, C. Cortes, N. Lawrence, and K. Q. Weinberger (Eds.), Vol. 27. Curran Associates, Inc. https://proceedings.neurips.cc/paper_files/paper/2014/ file/532a2f85b6977104bc93f8580abbb330-Paper.pdf [223] Le Yu, Bowen Yu, Haiyang Yu, Fei Huang, and Yongbin Li. 2024. Language models are super mario: absorbing abilities from homologous models as a free lunch. In Proceedings of the 41st International Conference on Machine Learning (ICML’24, Vol. 235). JMLR.org, Vienna, Austria, 57755–57775. [224] Qiying Yu, Zheng Zhang, Ruofei Zhu, Yufeng Yuan, Xiaochen Zuo, YuYue, Weinan Dai, Tiantian Fan, Gaohong Liu, Juncai Liu, LingJun Liu, Xin Liu, Haibin Lin, Zhiqi Lin, Bole Ma, Guangming Sheng, Yuxuan Tong, Chi Zhang, Mofan Zhang, Ru Zhang, Wang Zhang, Hang Zhu, Jinhua Zhu, Jiaze Chen, Jiangjie Chen, Chengyi Wang, Hongli Yu, Yuxuan Song, Xiangpeng Wei, Hao Zhou, Jingjing Liu, Wei-Ying Ma, Ya-Qin Zhang, Lin Yan, Yonghui Wu, and Mingxuan Wang. 2025. DAPO: An Open-Source LLM Reinforcement Learning System at Scale. https://openreview.net/forum? id=2a36EMSSTp&referrer=%5Bthe%20profile%20of%20YuYue%5D%28%2Fprofile%3Fid%3D~YuYue2%29 [225] Weizhe Yuan, Richard Yuanzhe Pang, Kyunghyun Cho, Xian Li, Sainbayar Sukhbaatar, Jing Xu, and Jason Weston. 2024. Self-rewarding language models. In Proceedings of the 41st International Conference on Machine Learning (ICML’24, Vol. 235). JMLR.org, Vienna, Austria, 57905–57923. [226] Eric Zelikman, Yuhuai Wu, Jesse Mu, and Noah D. Goodman. 2022. STaR: self-taught reasoner bootstrapping reasoning with reasoning. In Proceedings of the 36th International Conference on Neural Information Processing Systems (NIPS ’22). Curran Associates Inc., Red Hook, NY, USA, 15476–15488. [227] Binchi Zhang, Zhengzhang Chen, Zaiyi Zheng, Jundong Li, and Haifeng Chen. 2026. LOKA: Conflict-Aware LLM Knowledge Update with Adaptive Knowledge Memory. In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), Maria Liakata, Viviane P. Moreira, Jiajun Zhang, and David Jurgens (Eds.). Association for Computational Linguistics, San Diego, California, United States, 16689–16715. doi:10.18653/v1/2026.acl-long.760 [228] Qingru Zhang, Minshuo Chen, Alexander Bukharin, Nikos Karampatziakis, Pengcheng He, Yu Cheng, Weizhu Chen, and Tuo Zhao. 2023. AdaLoRA: Adaptive Budget Allocation for Parameter-Efficient Fine-Tuning. doi:10.48550/arXiv. 2303.10512 [229] Tianjun Zhang, Shishir G. Patil, Naman Jain, Sheng Shen, Matei Zaharia, Ion Stoica, and Joseph E. Gonzalez. 2024. RAFT: Adapting Language Model to Domain Specific RAG. In First Conference on Language Modeling. https: //openreview.net/forum?id=rzQGHXNReU [230] Chujie Zheng, Shixuan Liu, Mingze Li, Xiong-Hui Chen, Bowen Yu, Chang Gao, Kai Dang, Yuqiong Liu, Rui Men, An Yang, Jingren Zhou, and Junyang Lin. 2025. Group Sequence Policy Optimization. doi:10.48550/arXiv.2507.18071 [231] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric P. Xing, Hao Zhang, Joseph E. Gonzalez, and Ion Stoica. 2023. Judging LLM-as-a-judge with MT-bench and Chatbot Arena. In Proceedings of the 37th International Conference on Neural Information Processing Systems (NIPS ’23). Curran Associates Inc., Red Hook, NY, USA, 46595–46623.

42

Afdideh, Seoane, and Abtahi

[232] Zaiyi Zheng, Guanghui Min, Yaochen Zhu, Liang Wu, Liangjie Hong, Chen Chen, and Jundong Li. 2026. SAPO: Step-Aligned Policy Optimization for Reasoning-Based Generative Recommendation. doi:10.48550/arXiv.2605.17648 [233] Yue Zhou, Chenlu Guo, Xu Wang, Yi Chang, and Yuan Wu. 2024. A Survey on Data Augmentation in Large Model Era. doi:10.48550/arXiv.2401.15422 [234] Fuzhen Zhuang, Zhiyuan Qi, Keyu Duan, Dongbo Xi, Yongchun Zhu, Hengshu Zhu, Hui Xiong, and Qing He. 2021. A Comprehensive Survey on Transfer Learning. Proc. IEEE 109, 1 (Jan. 2021), 43–76. doi:10.1109/JPROC.2020.3004555 [235] Daniel M. Ziegler, Nisan Stiennon, Jeffrey Wu, Tom B. Brown, Alec Radford, Dario Amodei, Paul Christiano, and Geoffrey Irving. 2020. Fine-Tuning Language Models from Human Preferences. doi:10.48550/arXiv.1909.08593 [236] Luisa Zintgraf, Kyriacos Shiarli, Vitaly Kurin, Katja Hofmann, and Shimon Whiteson. 2019. Fast Context Adaptation via Meta-Learning. In Proceedings of the 36th International Conference on Machine Learning. PMLR, 7693–7702. https://proceedings.mlr.press/v97/zintgraf19a.html [237] Andy Zou, Long Phan, Sarah Chen, James Campbell, Phillip Guo, Richard Ren, Alexander Pan, Xuwang Yin, Mantas Mazeika, Ann-Kathrin Dombrowski, Shashwat Goel, Nathaniel Li, Michael J. Byun, Zifan Wang, Alex Mallen, Steven Basart, Sanmi Koyejo, Dawn Song, Matt Fredrikson, J. Zico Kolter, and Dan Hendrycks. 2025. Representation Engineering: A Top-Down Approach to AI Transparency. doi:10.48550/arXiv.2310.01405 [238] Andy Zou, Long Phan, Justin Wang, Derek Duenas, Maxwell Lin, Maksym Andriushchenko, Rowan Wang, Zico Kolter, Matt Fredrikson, and Dan Hendrycks. 2024. Improving Alignment and Robustness with Circuit Breakers. In Advances in Neural Information Processing Systems, A. Globerson, L. Mackey, D. Belgrave, A. Fan, U. Paquet, J. Tomczak, and C. Zhang (Eds.), Vol. 37. Curran Associates, Inc., 83345–83373. doi:10.52202/079017-2651

A Six-Dimensional Taxonomy of Post-Training Adaptation Techniques with Applications in AI Governance

A

43

Methodology and Corpus Flow

A.1

Multivocal Literature Review Approach

This study adopts a MLR methodology, integrating peer-reviewed academic literature with practitioneroriented grey literature to construct a comprehensive taxonomy of AI model adaptation techniques [68]. The MLR approach is warranted because AI adaptation is heavily driven by industry practice: emergent techniques frequently appear in framework documentation, developer blogs, and industry technical reports months or years before formal academic publication. Restricting the review to peer-reviewed sources alone would omit rapidly evolving concepts, such as CE, that have already achieved widespread practitioner adoption but lack formal academic treatment. The rationale for grey literature inclusion is developed in Section A.12.1. A.2 Search Strategy and Snowball Sampling Because terms such as FT, adaptation, and TL are heavily overloaded across the ML literature, standard database keyword searches yield fragmented and inconsistent results. We therefore adopted a purposive snowball sampling methodology, starting from canonical seed papers and tracing the semantic evolution of adaptation terminology through two complementary directions: backward snowballing, examining reference lists of seed papers to identify foundational works that established the core terminology and theoretical underpinnings of each adaptation paradigm; and forward snowballing, identifying citing papers via Google Scholar, Semantic Scholar, and arXiv to discover emergent variants, extensions, and successor methods that refine or supersede the seed techniques. Twelve academic snowballing chains were traced, each anchored by a canonical seed paper or survey cluster: (1) PEFT ([90]); (2) TL ([153] and [234]); (3) Model Merging and Task Arithmetic ([100]); (4) RL from Verifiable Rewards ([74] and [175]); (5) Alignment Methods ([164] and [152]); (6) CL ([178]); (7) Meta-Learning ([63] and [87]); (8) RAG ([114]); (9) Knowledge Editing ([140]); (10) Machine Unlearning ([36]); (11) Representation Engineering (RepE) and Activation Steering ([237] and [196]); (12) Data Augmentation and Synthetic Data Generation ([205] and [233]). A.3

Inclusion and Exclusion Criteria

Sources were included if they met at least one of the following criteria: (a) they introduced, formalized, or significantly extended an adaptation technique applicable to ML models; (b) they provided a systematic survey or taxonomic organization of adaptation methods; or (c) they constituted authoritative framework documentation or technical reports from major AI organizations describing adaptation tooling or methodology. Sources were excluded if they addressed foundational learning paradigms (e.g., supervised, unsupervised, or reinforcement learning) without an adaptation context, or if they focused exclusively on low-level optimization mechanics (e.g., GD algorithms, learning rate scheduling) rather than adaptation strategies.

44

A.4

Afdideh, Seoane, and Abtahi

Grey Literature

To capture practitioner terminology and tooling that precedes formal publication, the review incorporated four categories of grey literature: • Framework documentation: official documentation for Hugging Face PEFT (v0.18.0), MergeKit (Arcee AI), Transformer Reinforcement Learning (TRL), ms-swift (ModelScope), and EasyEdit (Zhejiang University); • Developer blogs and technical reports from Anthropic, LangChain, LlamaIndex, Weaviate, and the PE Guide describing emergent techniques and best practices; • Industry analysis, including reports from organizations such as Gartner tracking the adoption trajectory of adaptation terminology (e.g., the shift from PE to CE); • Synthetic data and augmentation surveys, comprising preprints and workshop papers covering data-centric adaptation methods that bridge academic and industrial practice. A.5

Corpus and Synthesis

The corpus was assembled through a staged process: seed identification, backward snowballing, forward snowballing, and grey literature integration. Academic sources include foundational papers, tier-1 venue publications (NeurIPS, ICML, ICLR, ACL, IEEE TPAMI, ACM Computing Surveys), and high-impact arXiv preprints with substantial citation counts. Grey literature sources were included only if they met at least one of three criteria: (a) they constituted official framework documentation from a major AI organization; (b) they introduced or formalized a technique or term not yet covered in peer-reviewed literature; or (c) they provided implementation-level detail necessary to distinguish technique variants that appear identical in academic descriptions. Promotional material, opinion pieces, and sources without verifiable technical content were excluded. A.6

Saturation Criterion and Stopping Rule

Two distinct criteria governed corpus growth. A candidate required a new row when its D1 – D6 profile, i.e., the combination of category values across all six dimensions, was not already represented, and it constituted an independently selectable adaptation decision unit. A candidate required a new dimension category when it could not be characterized by any existing value on some dimension. Saturation was defined as the point at which no further candidate required either. The category structure was developed jointly with the corpus rather than fixed in advance, and was extended where the literature demanded. Several category values are instantiated solely by techniques traced through a single chain and entered the framework with that chain’s integration: d1 = Activation-Space through the activation-steering chain (Row 49), as direct algebraic intervention on intermediate representations is structurally distinct from gradient-based parametric update; d2 = Reasoning and d2 = Verifiable Task through the RLVR/GRPO chain (Row 16), whose seed literature also motivated d1 = Inference-Time Search (Row 42); and d2 = Remove Knowledge with d3 = Forget Set Specification through the machine-unlearning chain (Row 47). The final 48-technique set satisfies both criteria: every candidate identified through forward snowballing and grey-literature review, including those integrated late in the process, was either assignable to an existing row or accommodated by a row differing from all existing rows on at least one dimension, without requiring further extension of the category structure. Candidate paradigms that were evaluated and not added, including agentic orchestration patterns, were determined to represent compositions of existing techniques or open research questions rather than distinct adaptation techniques. Chains are presented in Section A.2 in thematic order; integration order differed and is not load-bearing for this criterion. The 48-technique count is the saturation count under the D1 –D6 criterion, not a claim that no further techniques exist. Techniques whose D1 –D6 profile is shared with an existing row are subsumed

A Six-Dimensional Taxonomy of Post-Training Adaptation Techniques with Applications in AI Governance

45

under that row even when they carry distinct names in the literature. New rows are required only when a candidate presents a profile no existing row provides. A.7

Classification Process

The synthesis proceeded iteratively: individual technique definitions were first extracted and crossreferenced across sources to resolve terminological inconsistencies, then organized into the multidimensional characterization framework. Technique classifications were drafted by the first author and independently reviewed by all authors against the D1 –D6 category definitions; disagreements were resolved by discussion until consensus. Classification tensions, that is, cases where techniques resist clean single-category assignment on one or more dimensions, were identified during this process and are documented transparently in Section C.10. Taxonomy gaps, defined as areas where the reviewed literature revealed adaptation paradigms not yet integrated into existing frameworks, were identified through systematic comparison of the snowballing chains against the emerging taxonomy structure. Five such gaps were identified: RLVR/GRPO, machine unlearning, knowledge editing, CE, and activation steering. All five were subsequently resolved and integrated into the final framework. A.8

Literature Trace by Snowballing Chain

The purposive snowball sampling strategy described in Section A.2, combined with the greyliterature rationale of Section A.1, produced twelve independent chains. Tables 11–22 report the full trace for each chain, separating backward-snowballing foundations, seed papers, and forwardsnowballing successors, with the key development each source contributed. A.9

Search and Selection Process

The corpus was assembled through the following staged process, summarized in Section A.10: (1) Seed identification. 12 snowball chains were defined, each corresponding to a distinct adaptation paradigm (Tables 11–22). Each chain was anchored by one or two canonical seed papers selected for citation impact and foundational status; where a paradigm required both a seminal method paper and a comprehensive survey, two co-seeds were used. This yielded 18 distinct seed papers across 12 chains. (2) Backward snowballing. For each seed, reference lists were examined to identify foundational works that established the core terminology and theoretical underpinnings of each adaptation paradigm. All 12 chains received backward snowballing; foundational references are listed per-chain in Tables 11–22. This yielded 36 foundational sources across all chains. (3) Forward snowballing. Citing papers were traced via Google Scholar, Semantic Scholar, and arXiv to discover emergent variants, extensions, and successor methods. This yielded 69 additional academic sources. (4) Grey literature integration. Framework documentation, developer blogs, technical reports, and industry analyses were incorporated where they captured practitioner terminology or tooling that precedes or complements formal publication. This yielded 20 grey literature sources (see Section A.12). (5) Exclusion. Sources were excluded if they: (a) addressed foundational learning paradigms without an adaptation context; (b) focused exclusively on low-level optimization mechanics; or (c) constituted promotional material without technical substance. (6) Final corpus. Searches were conducted through July 2026. The resulting corpus comprises 123 academic sources, spanning 1985–2026, and 20 grey literature sources. Academic sources include foundational papers, tier-1 venue publications (NeurIPS, ICML, ICLR, ACL, IEEE

46

Afdideh, Seoane, and Abtahi

TPAMI, ACM Computing Surveys), and high-impact arXiv preprints with substantial citation counts. Dating convention and apparent temporal inversions. Snowball chains operate on the citation graph, which is driven by preprint availability rather than venue publication dates. Because of preprint-to-publication lag (typically 6–18 months for NeurIPS/ICML/ICLR, and up to two years for journal venues such as IEEE TPAMI or ACM Computing Surveys), a paper can legitimately appear in the reference list of a work whose venue year precedes or equals its own. For example, DeepSeekMath (2024, [175]) cites Let’s Verify Step by Step (ICLR 2024, [123]) via its May 2023 preprint. Throughout this appendix, chain chronology follows first-preprint dates, while tables report its publication year. A.10

Corpus Flow Summary

While this review follows an MLR rather than strict systematic literature review protocol, the following flow summarizes the corpus assembly process: • Snowball chains: 12 (Tables 11–22) • Seed: 18 sources • Backward snowballing: 36 sources • Forward snowballing: 69 sources • Grey literature: 20 sources • Final corpus: 123 academic sources + 20 grey literature sources • Techniques extracted: 48 distinct adaptation techniques (Rows 2–49 of the centerpiece table; Row 1, Training, is a baseline reference, not a post-training technique) A.11

Snowballing Results by Adaptation Paradigm

This section presents the results of our snowball sampling methodology, organized across 12 tables (Tables 11–22) that correspond to the core adaptation paradigms evaluated in this study. Each table maps the methodological evolution of a specific paradigm, structured around three distinct phases of our literature trace: (1) the foundational works identified through backward snowballing; (2) the canonical seed papers selected to anchor the chain; and (3) the emergent variants, extensions, and successor methods discovered via forward snowballing. To contextualize each source, every entry provides the method’s name, its reference, and a concise summary of the key development it contributed to the post-training adaptation landscape.

A Six-Dimensional Taxonomy of Post-Training Adaptation Techniques with Applications in AI Governance

Table 11. PEFT snowballing results. Method

Ref.

Key Development I. Backward Snowballing

—

[15]

Shows that the success of FT LLMs on small datasets is due to their extremely low intrinsic dimension and that larger models inherently possess lower intrinsic dimensions. Enable highly parameter-efficient transfer learning by injecting a small number of trainable, task-specific parameters into a frozen pre-trained model.

Adapter Modules

[88]

LoRA

[90]

Reduces the computational cost and memory requirements of FT LLMs by freezing pre-trained weights and injecting small, trainable rank decomposition matrices.

DoRA

[127]

AdaLoRA

[228]

VeRA

[110]

QLoRA

[52]

rsLoRA

[104]

LoRA+

[83]

PiSSA

[139]

LoRA-Pro DuDe

[210] [80]

FedPara

[98]

DyLoRA

[198]

MoELoRA

[132]

IA3

[125] [203]

Bridges the accuracy gap between LoRA and full FT by decomposing pre-trained weights into magnitude and direction, using LoRA solely for efficient directional updates to improve learning capacity without adding inference overhead. Dynamically allocates the parameter budget across weight matrices based on their importance, utilizing SVD to efficiently prune less critical updates. Dramatically reduces the parameter and storage footprint of LoRA by sharing a single, frozen pair of random low-rank matrices across all layers and learning only small, layer-specific scaling vectors. Enables the FT of LLMs (up to 65B parameters) on a single 48GB GPU by combining a frozen, 4-bit quantized base model with LoRA, maintaining full 16-bit performance via memory-saving innovations like 4-bit NormalFloat, Double Quantization, and Paged Optimizers. Modifies the traditional scaling factor, thereby unlocking the effective use of higher-rank adapters to achieve better FT performance. Improves the FT speed and performance of large-width models by assigning different learning rates to the adapter matrices A and B. Accelerates PEFT by initializing adapter matrices with the original model’s principal singular components and freezing the residual weights, resulting in faster convergence, reduced quantization error, and superior task performance. Adjusts the gradients of LoRA’s low-rank matrices to more accurately approximate the full FT gradient. Decomposes weight matrices into magnitude and direction components and utilizes SVD for principled initialization. A communication-efficient parameterization technique for FL that utilizes a low-rank Hadamard product to dramatically reduce model upload and download costs without sacrificing model capacity, while also offering an extension for highly efficient personalized FL. Dynamically trains adapter modules across a range of ranks simultaneously rather than a single fixed rank, eliminating the need for exhaustive rank search and enabling flexible rank adjustments post-training without the need to retrain from scratch. Treats LoRA modules as a MoE and employs contrastive learning to encourage distinct feature learning among these experts. Scales internal activations by learned vectors, and pairs it with the T-Few recipe. Improves FT by restricting updates to the minor singular components of the weight matrix, which represent noisy or long-tail information, while keeping the principal components frozen to preserve essential pre-trained knowledge.

II. Seed

III. Forward Snowballing

MiLoRA

47

48

Afdideh, Seoane, and Abtahi

Table 12. TL snowballing results. Method

Ref.

Key Development I. Backward Snowballing

MTL

[37]

Mixture Model

[99]

—

[179]

Self-Taught Learning —

[165]

BERT

[53]

—

[153]

—

[234]

—

[133]

—

[75]

—

[30]

—

[204]

[222]

Improves model generalization by simultaneously training related tasks on a shared representation, allowing the domain information from one task to act as a beneficial inductive bias for the others. Uses conditional expectation maximization to adapt maximum entropy classifiers to new domains, effectively leveraging abundant out-of-domain training data alongside scarce in-domain data to improve predictive performance. Improves predictive inference under covariate shift by weighting the log-likelihood function using the ratio of population to observation covariate densities, demonstrating asymptotic optimality under model misspecification and introducing an information criterion to select optimal weights for moderate sample sizes. Leverages vast amounts of random, unstructured unlabeled data using sparse coding to construct higher-level feature representations that significantly improve performance on supervised classification tasks. Experimentally quantifies feature transferability across deep neural network layers, revealing that while transfer is hindered by higher-layer specialization and the splitting of co-adapted neurons, initializing networks with transferred features yields a lasting generalization boost even across distant tasks. Pre-trains deep bidirectional representations that condition on both left and right context across all layers. II. Seed Formalizes how knowledge can be effectively transferred across different feature spaces and data distributions to overcome the standard ML assumption of identical training and testing distributions. Reviews over 40 representative approaches from data and model perspectives and provides an extensive empirical evaluation of more than 20 models across multiple datasets to guide practical model selection. III. Forward Snowballing Measures differences in internal dependence structures separately from marginal distributions, optimizing their relative weights to robustly enhance transferability where domain shifts matter most. Reviews FTL, outlining how it solves challenges like data heterogeneity and scarcity while preserving privacy constraints across decentralized nodes. Introduces a distribution-free, kernel-based framework that utilizes reproducing kernel Hilbert spaces and regularized empirical risk minimization to achieve universal consistency when generalizing classification labels from multiple related training datasets to a new, unlabeled sample. Reviews theoretical foundations and categorizes existing algorithms into data manipulation, representation learning, and learning strategies to facilitate the development of models that maintain robust performance when deployed in unseen, out-of-distribution environments.

Table 13. Model merging snowballing results. Method

Ref.

Key Development I. Backward Snowballing

Model Soups

[214]

Fisher Merging SWA

[135]

Task Arithmetic

[100]

A paradigm for model editing that represents task-specific behavioral shifts as "task vectors" (the difference between fine-tuned and pre-trained model weights) which can be intuitively added, negated, or combined to steer model performance across multiple tasks without additional training.

TIESMERGING

[218]

DARE

[223]

Model Breadcrumbs

[49]

DELLAMerging, MAGPRUNE EMRMerging SLERP

[51]

Enables the effective combination of multiple task-specific models into a single multitask model by mitigating parameter interference through a three-step process: pruning redundant parameter changes, resolving crossmodel sign conflicts, and exclusively merging parameters that align with the agreed-upon sign. Eliminates the vast majority of redundant FT deltas to mitigate interference, enabling multiple task-specific language models to be fused into a single, high-performing model without the need for retraining. Creates sparse, task-specific weight updates (derived from the delta between pre-trained and fine-tuned weights) to enable efficient, hyperparameter-free integration of multiple task adaptations into a single foundation model. Assigns higher dropout probabilities to parameters with smaller values and applies a compensatory rescaling factor to preserve embedding integrity.

LATA

[41]

[101]

Improves model accuracy and robustness by averaging the weights of multiple models fine-tuned with different hyperparameter configurations, without increasing inference or memory costs. Combines models by computing a weighted average of their parameters based on their Fisher information matrices. Improves generalization and finds flatter loss optima by averaging multiple weights sampled along the SGD training trajectory. II. Seed

III. Forward Snowballing

[95] [180]

Aligns multiple task-specific models with a unified base model by generating lightweight, task-specific masks and rescalers, thereby enabling multi-task capabilities without requiring additional data or training. Interpolates between compatible model checkpoints along aspherical path. The method originates in spherical interpolation for rotations and is implemented in contemporary LLM model-merging toolchains, including MergeKit. It has also been evaluated empirically as a model-merging strategy [131, 180]. Assigns layer-specific weights to task vectors, isolating task-specific knowledge from shared instructionfollowing components.

A Six-Dimensional Taxonomy of Post-Training Adaptation Techniques with Applications in AI Governance

Table 14. RLVR/GRPO snowballing results. Method

Ref.

Key Development I. Backward Snowballing

— InstructGPT PPO —

[43] [152] [170] [123]

Enables complex goal specification through human preference feedback between trajectory segments. Aligns LLMs with human intent by using SFT on demonstrations followed by RLHF. A surrogate objective function that facilitates multiple epochs of minibatch updates. Demonstrates how providing human feedback at each intermediate step of a logical chain outperforms feedback provided only at the final result for training reliable LLMs on complex mathematical reasoning tasks.

DeepSeekMath, [175] GRPO DeepSeek-R1 [74]

Utilizes a curated pipeline for high-quality web data and introducing GRPO, a memory-efficient alternative to PPO that optimizes reasoning performance. Incentivizes advanced reasoning capabilities, such as self-reflection and dynamic verification, directly in LLMs without relying on human-annotated reasoning trajectories.

DAPO

[224]

GSPO

[230]

SAPO

[232]

CISPO RLOO

[144] [109]

Enhances long-chain-of-thought RL through four technical innovations (Clip-Higher, Dynamic Sampling, Token-Level Policy Gradient Loss, and Overlong Reward Shaping) designed to improve training stability, efficiency, and reasoning performance. Shifts from traditional token-level importance ratios to sequence-level optimization to achieve enhanced stability and performance, particularly for MoE architectures. Treats individual reasoning steps as the fundamental unit for credit assignment, which stabilizes training and improves performance. Enhances training efficiency by clipping importance sampling weights rather than token updates. Introduces a variance-reducing REINFORCE estimator that draws multiple samples without replacement per datapoint to establish a natural baseline. a critic-free framework that utilizes global advantage normalization across the entire batch rather than local prompt-specific groups, providing an effectively unbiased and stable estimate.

II. Seed

III. Forward Snowballing

REINFORCE++ [92]

Table 15. Alignment snowballing results. Method

Ref.

Key Development I. Backward Snowballing

—

[235]

Introduces a framework for applying RL to natural language tasks by training a reward model based on human preference comparisons, demonstrating effective performance on stylistic continuation and summarization tasks.

InstructGPT DPO

[152] [164]

Aligns LLMs with human intent by using SFT on demonstrations followed by RLHF. Mathematically bypasses the complex, unstable RLHF pipeline by mapping reward functions directly to optimal policies, allowing for stable, single-stage fine-tuning based on human preference classification.

DeepSeekMath, [175] GRPO ORPO [86]

Utilizes a curated pipeline for high-quality web data and introducing GRPO, a memory-efficient alternative to PPO that optimizes reasoning performance. Eliminates the need for a separate preference alignment phase by embedding preference optimization directly into SFT using a minor odds ratio penalty for the disfavored style. A reference-free offline preference optimization algorithm that uses the average sequence log probability as an implicit reward and incorporates a target reward margin. An alignment objective rooted in prospect theory that directly maximizes the human utility of model generations, matching or exceeding preference-based methods using only a simple binary signal of desirability. Improves machine translation by teaching models to avoid merely adequate but imperfect translations. An offline preference optimization algorithm inspired by Mallows’ ranking theory that incorporates a dispersion index to model the diversity of human preferences. Utilizes four key techniques to scale RL and elicit complex reasoning in large language models. A general objective that bypasses the pointwise reward approximations by learning directly from pairwise preferences.

II. Seed

III. Forward Snowballing

SimPO

[142]

KTO

[58]

CPO MallowsPO

[217] [39]

DAPO IPO

[224] [22]

49

50

Afdideh, Seoane, and Abtahi

Table 16. CL snowballing results. Method

Ref.

Key Development I. Backward Snowballing

Catastrophic Interference EWC

[136]

—

[27]

—

[178]

A survey, defining the landscape through "vertical" (general-to-specific) and "horizontal" (across time/domains) continuity, and structuring research into three distinct stages: Continual Pre-Training, Domain-Adaptive Pre-training, and Continual FT.

—

[206]

A survey that establishes a taxonomy of methods, clarifies the stability-plasticity trade-off, and evaluates how various strategies (such as regularization, replay, and architecture-based approaches) manage the stabilityplasticity trade-off.

[108]

Sequential learning in connectionist networks causes new information to destructively overwrite previously learned representations. This phenomenon explains the catastrophic forgetting observed in modern full FT. A method inspired by synaptic consolidation in neuroscience that overcomes catastrophic forgetting by identifying and selectively slowing down updates to weights that are critical for previously learned tasks. Compares various gradient-based training algorithms and activation functions, identifying dropout as the regularization technique that consistently provides the best stability–plasticity trade-off. II. Seed

III. Forward Snowballing

Table 17. Meta-learning snowballing results. Method

Ref.

Key Development I. Backward Snowballing

MetaLearning Matching Networks Prototypical Networks

[193]

MAML

[63]

—

[87]

MAML++

[18]

FOMAML, Reptile

[151]

Meta-SGD

[122]

[201] [185]

Enables machines to improve their own generalization capabilities through experience, mirroring the human ability to acquire "bias" and learn effectively from extremely sparse data. Combines metric learning with external memory to enable one-shot learning by directly predicting labels from a support set of examples without requiring task-specific FT. Simplifies few-shot classification by learning a metric space where new examples are classified based on their distance to class-specific prototype representations, demonstrating that a straightforward inductive bias can outperform complex meta-learning architectures. II. Seed An optimization-based meta-learning algorithm that explicitly trains model parameters to be highly sensitive to gradient updates, allowing a model to achieve strong generalization on new tasks after only a few training samples and minimal gradient steps. This survey establishes a comprehensive meta-learning taxonomy and evaluates the field’s progress in overcoming deep learning bottlenecks, such as data scarcity and generalization challenges, by enabling algorithms to improve themselves through experience across multiple tasks. III. Forward Snowballing

LEO

[168]

CAVIA

[236]

Improvements to the MAML framework that resolves its inherent training instability, hyperparameter sensitivity, and computational overhead while significantly improving generalization and convergence speed. Explicitly optimizes a model’s initial parameters so that produces rapid adaptation and high generalization across diverse learning tasks. An efficient FOMAML that learns an adaptable parameter initialization by repeatedly moving the initial weights toward the weights learned on sampled tasks, eliminating the need for computationally expensive second-order derivatives. Extends upon MAML by simultaneously learning a model’s optimal initialization, update directions, and learning rates to enable highly efficient, one-step adaptation to new tasks. Bypasses the limitations of high-dimensional parameter spaces in FSL by learning a data-dependent generative representation and performing gradient-based meta-adaptation entirely within a low-dimensional latent space. Improves upon MAML by partitioning model parameters into globally shared weights and task-specific context parameters, restricting test-time adaptation solely to the low-dimensional context to reduce overfitting and improve efficiency.

A Six-Dimensional Taxonomy of Post-Training Adaptation Techniques with Applications in AI Governance

Table 18. RAG snowballing results. Method

Ref.

Key Development

DPR

[105]

A method for open-domain question answering that replaces traditional sparse retrieval models (like TF-IDF or BM25) with dense embeddings learned via a dual-encoder framework. Presents the PRF conceptually, incorporating document metadata, such as structure and link-graph information, culminating in the BM25F algorithm for highly effective web and corporate search retrieval. Achieves performance comparable to massive models like GPT-3 with 25 times fewer parameters by augmenting an auto-regressive language model with a frozen BERT retriever and chunked cross-attention to condition predictions on document chunks retrieved from a 2-trillion-token database.

I. Backward Snowballing

PRF

[167]

RETRO

[32]

RAG

[114]

Combines pre-trained parametric memory (a seq2seq model) with non-parametric memory (a dense vector index accessed via a neural retriever).

— Agentic RAG

[67] [183]

Self-RAG

[19]

GraphRAG

[57]

HippoRAG

[78]

RAFT

[229]

A comprehensive review of the evolution of RAG paradigms from Naive to Modular. A comprehensive analytical survey of Agentic RAG, introducing a principled taxonomy that details how embedding autonomous agents into RAG pipelines overcomes static workflow limitations through dynamic retrieval, multi-step reasoning, and adaptive collaboration. Trains a language model to adaptively retrieve information on-demand and use special reflection tokens to evaluate both the retrieved passages and its own generations, improving factuality and inference-time controllability. A graph-based RAG framework that addresses the limitations of standard RAG on global, corpus-wide queries by constructing an entity knowledge graph and pregenerating community summaries that are aggregated to synthesize comprehensive responses across massive text collections. Inspired by human long-term memory that orchestrates LLMs and knowledge graphs to enable efficient, single-step knowledge integration across new experiences. A post-training recipe that enhances in-domain, open-book question answering by teaching models to ignore irrelevant distractor documents and cite verbatim passages from relevant texts using chain-of-thought reasoning.

II. Seed

III. Forward Snowballing

Table 19. Knowledge editing snowballing results. Method

Ref.

Key Development I. Backward Snowballing

Knowledge [47] Neurons KnowledgeEditor [50]

Identifies the specific neurons responsible for storing and expressing factual knowledge within pretrained Transformers, enabling direct knowledge editing without model FT. Corrects specific incorrect or obsolete factual knowledge in language models without expensive retraining by utilizing a hyper-network trained with constrained optimization to predict precise weight updates that generalize to paraphrases while leaving unrelated knowledge intact. Reveals that feed-forward layers in transformer language models operate as interpretable key-value memories, where keys detect specific textual patterns and values induce corresponding output vocabulary distributions to predict subsequent tokens.

—

[69]

ROME

[140]

Uses causal interventions to locate factual associations within the mid-layer feed-forward modules of autoregressive transformers and directly updates these specific weights to alter factual predictions without sacrificing specificity or generalization.

MEMIT MEND

[141] [145]

SERAC PMET

[146] [119]

WISE

[207]

MAKE

[155]

A scalable method for directly updating LLMs that enables the simultaneous insertion of thousands of memories. An efficient post-hoc editing approach that utilizes small auxiliary networks and low-rank gradient decomposition to enable fast, targeted corrections to massive pre-trained models using just a single input-output pair. Stores edits in an explicit memory and reasons over them to dynamically modulate the base model’s predictions. Simultaneously optimizes Transformer component hidden states but selectively updates only the Feed-Forward Network weights, thereby preserves the general knowledge extraction patterns inherently encoded within the Multi-Head Self-Attention mechanism. A dual parametric memory scheme that resolves the impossible triangle of lifelong model editing by employing a router to direct queries between a main pre-trained memory and a side memory, where continual edits are seamlessly integrated via a conflict-free knowledge-sharding mechanism. Leverages a language model’s own internal memory to recall and update indirectly associated knowledge, ensuring consistent knowledge transfer without the need for external databases.

II. Seed

III. Forward Snowballing

51

52

Afdideh, Seoane, and Abtahi

Table 20. Machine unlearning snowballing results. Method

Ref.

Key Development I. Backward Snowballing

Machine Unlearning

[71]

To comply with privacy mandates like the “Right to Be Forgotten”, it proposes algorithmic principles and specific 𝑘 -means algorithms that efficiently delete individual data points without the prohibitive computational cost of retraining models from scratch. Accelerates machine unlearning by strategically partitioning data and isolating its influence during training, significantly reducing the computational overhead required to remove a user’s data from deep neural networks compared to retraining from scratch.

SISA

[33]

Machine Unlearning

[36]

Enables systems to rapidly and completely forget specific training data and its lineage by transforming learning algorithms into a summation form, achieving data removal asymptotically faster than retraining from scratch.

GA

[220]

Task Vec. Negation RMU

[100]

Adaptive RMU

[97]

SAEs LoKU, FILA

[61] [38]

LOKA

[227]

Maximize loss on forget set (reverse learning signal) Demonstrates that integrating GA with GD on indistribution data provides a robust, highly efficient mechanism to enforce the "right to be forgotten" without the computational burden of retraining. A paradigm for steering neural network behavior by representing tasks as vectors in weight space that can be added, negated, or combined to efficiently edit models and improve multi-task performance. Introduces the WMDP benchmark to measure hazardous knowledge in LLMs, alongside RMU, an unlearning method that controls model representations to selectively erase dangerous capabilities while preserving general domain proficiency. Provides a theoretical framework explaining how RMU successfully erases knowledge by reducing token confidence, and introduces Adaptive RMU, an enhancement that overcomes standard RMU’s limitations to enable highly effective unlearning across all network layers. Demonstrates that negatively scaling specific interpretable features can selectively erase targeted knowledge. Introduces LoKU, a framework that leverages an Inverted Hinge Loss to suppress sensitive tokens while maintaining linguistic fluency and utilizes FILA to precisely target unlearning without sacrificing a model’s broader reasoning and generative capabilities. A conflict-aware framework for simultaneous knowledge unlearning and updating that utilizes an adaptive, multi-unit memory architecture and a learning-based router to dynamically integrate new information while effectively managing task conflicts.

II. Seed

III. Forward Snowballing

[118]

Table 21. RepE / Activation steering snowballing results. Method

Ref.

Key Development I. Backward Snowballing

—

[143]

TCAV

[107]

ITI

[117]

RepE

[237]

ActAdd

[196]

CAA FVs

[166] [194]

Conceptor

[160]

Circuit Breakers

[238]

Demonstrates that vector-space word representations learned by continuous language models implicitly capture deep syntactic and semantic regularities, enabling complex linguistic relationships to be accurately modeled and solved through simple algebraic vector operations—famously illustrated by "King - Man + Woman = Queen." Translate a neural network’s internal representations into high-level, human-interpretable concepts and use directional derivatives to quantify their exact importance to classification decisions. Enhances the truthfulness of LLMs by shifting internal activations within specific attention heads during inference, offering a computationally inexpensive and data-efficient method to surface internally represented truths. II. Seed A neuroscience-inspired, top-down approach to AI transparency that focuses on monitoring and manipulating population-level representations rather than individual neurons to effectively control safety-relevant behaviors like honesty and harmlessness in LLMs. Steers language model outputs by directly modifying intermediate activations during inference using simple steering vectors derived from prompt pairs, enabling real-time control over high-level properties like sentiment and toxicity without the need for FT. III. Forward Snowballing Steers the behavior of LLMs at inference time by modifying their internal hidden state activations. Demonstrates that LLMs encode specific tasks as compact, extractable FVs within their attention heads, which can be independently manipulated or combined to causally steer the model’s behavior across novel contexts. Introduces “conceptors” for activation engineering inference-time steering of LLMs and the use of Boolean operations to effectively combine complex control goals. Introduces “circuit breakers,” an alignment technique that directly disrupts the internal model representations responsible for harmful outputs, creating a robust, attack-agnostic defense for both text and multimodal AI systems without compromising their core utility.

A Six-Dimensional Taxonomy of Post-Training Adaptation Techniques with Applications in AI Governance

Table 22. Data augmentation and synthetic data generation snowballing results. Method

Ref.

Key Development I. Backward Snowballing

—

[173]

Introduces a back-translation strategy for NMT that pairs target-side monolingual data with automatic synthetic translations to create additional parallel training data. A technique consisting of four simple operations (synonym replacement, random insertion, random swap, and random deletion) that boosts text classification performance and data efficiency, particularly on smaller datasets. Automates the creation of data augmentation strategies by using a search algorithm to discover optimal, highly transferable policies composed of varied image-processing sub-policies that maximize a neural network’s validation accuracy.

EDA

[212]

AutoAugment

[45]

— —

[205] [233]

Provides a survey of data synthesis and augmentation techniques across the entire lifecycle of LLMs. This survey systematically categorizes and evaluates large model-driven data augmentation techniques across text, image, and paired data, highlighting how leveraging these models to generate high-quality training data addresses the critical challenge of impending data scarcity across multiple domains.

—

[128]

Systematizes the workflow of LLM-driven synthetic data generation, curation, and evaluation.

II. Seed

III. Forward Snowballing

53

54

A.12

Afdideh, Seoane, and Abtahi

Grey Literature Sources

A.12.1 Rationale for Grey Literature Inclusion. The inclusion of grey literature in this MLR is warranted on three grounds. First, AI adaptation is heavily driven by industry practice: emergent techniques frequently appear in framework documentation, developer blogs, and technical reports months or years before formal academic publication. Restricting the review to peer-reviewed sources alone would omit rapidly evolving concepts. Second, several key techniques in the framework (notably CE, model merging via SLERP, and the GRPO family algorithms) were first formalized in practitioner documentation and only subsequently (if at all) received academic treatment. Third, framework documentation (Hugging Face PEFT, MergeKit, TRL, ms-swift, EasyEdit) constitutes the de facto standard reference for implementation details and method comparisons that academic papers often omit or simplify. Grey literature sources were included only if they met at least one of the following criteria: (a) they constituted official framework documentation from a major AI organization; (b) they introduced or formalized a technique or term not yet covered in peer-reviewed literature; or (c) they provided implementation-level detail necessary to distinguish technique variants that appear identical in academic descriptions. Promotional material, opinion pieces, and sources without verifiable technical content were excluded. A.12.2 Grey Literature Inventory. The following is a complete enumeration of the 20 grey literature sources retained in the final corpus, organized by category: Framework documentation (5): Hugging Face PEFT (v0.18.0); MergeKit (Arcee AI); TRL (Hugging Face); ms-swift (ModelScope); EasyEdit (Zhejiang University, ACL 2024 System Demo). CE (6): Anthropic engineering blog; Prompt Engineering Guide; LlamaIndex guide; Weaviate guide; Philipp Schmid blog; Kubiya 2025 guide. See Section A.12.3. Industry technical reports (5): A. Karpathy (CE definition; LLM-as-CPU mental model); Gartner CE tracking; OpenAI system card documentation (RLHF pipeline details); DeepSeek-R1 technical report appendices (training pipeline implementation details beyond the main paper). Developer blogs and analyses (4): T. Lütke (Shopify) CE remarks; D. Horthy (HumanLayer) CE remarks; SLERP implementation guides (multiple community sources, consolidated); MergeKit community benchmarking reports. A.12.3 CE (Grey Literature). Definition: A. Karpathy: “The delicate art and science of filling the context window with just the right information for the next step” [17]. Distinction from PE: PE concerns how to phrase the question (static, single instruction). CE concerns designing the system architecture that feeds the LLM the right information at the right time (dynamic system, runtime assembly) [169]. Components: System prompt, user input, short-term memory (chat history), long-term memory (vector store retrieval), information from knowledge bases (RAG, MCP, API calls), tool definitions/responses, structured outputs, global state/workflow context [161]. Inclusion justification: Despite the absence of peer-reviewed formalization, CE is included in the framework (Row 40) because: (a) it has achieved widespread practitioner adoption, as evidenced by Gartner’s tracking of the term [6, 8]; (b) it occupies a distinct six-dimensional profile ( d1 = Context Injection , d3 = {External Corpus, User Interaction Data} , d4 = Version-Persistent , d5 = Input/Output-Space ) that is not captured by any other row; and (c) it represents the operational pattern underlying agentic AI systems [4, 5, 7]. Excluding it would leave a significant gap between the taxonomy and current practice.

A Six-Dimensional Taxonomy of Post-Training Adaptation Techniques with Applications in AI Governance

55

A.12.4 Hugging Face PEFT Library. Supported methods: LoRA, Prefix Tuning, Prompt Tuning, P-Tuning, AdaLoRA, LoHa, LoKr, AdaptionPrompt, IA3 , DoRA, VeRA, Context-Aware Prompt Tuning (CPT), CorDA, BitFit, Hotswapping (15+ methods) [11]. A.12.5 MergeKit (Arcee AI). Supported methods: Task Arithmetic, TIES-MERGING, DARE, SLERP, Model Soups, linear interpolation [9]. A.12.6 TRL Library (Hugging Face). Supported methods: SFT, DPO, KTO, Reward Modeling, CPO, SimPO, ORPO, PPO, REINFORCE [12]. A.12.7 ms-swift (ModelScope). Supported RLVR algorithms: GRPO, DAPO, GSPO, SAPO, CISPO, CHORD, RLOO, Reinforce++. Also supports: DPO, KTO, RM, CPO, SimPO, ORPO [10]. A.12.8 EasyEdit (Zhejiang University). Supported methods: ROME, MEMIT, MEND, SERAC, PMET, GRACE, WISE, EMMET, InstructEdit, MELO [13]. A.13

Snowball Saturation and Seed-Sensitivity Simulation

To evaluate the robustness of the purposive snowball sampling methodology and quantify the structural reliance on any single seed paper, we conducted a Monte Carlo (MC) sensitivity analysis (5,000 replicates). It is important to note that this simulation models the opportunity structure of the literature network rather than strictly reconstructing the historical citation log. It tests whether the 18 selected seeds were mathematically sufficient to reach the final retained corpus (123 chain entries across the 12 chains plus 20 grey-literature sources; Section A.10) A.13.1 Methodological Assumptions. Reachability was modeled using fixed category-level affinity probabilities rather than deterministic links: • Own-chain affinity (the probability of discovering an item within the same taxonomy category as the seed) was set to 0.90. • Related-chain affinity (cross-discovery between heavily overlapping domains, such as RLVR and Alignment, or Merging and Unlearning via task vectors) was parameterized between 0.20 and 0.45, mirroring the cross-chain duplicates observed during the actual review. • Unrelated-chain affinity was set to a baseline noise level of 0.03. • Grey literature accessibility was modeled at 0.18 from modern, LLM-era "practitioneradjacent" seeds, and 0.03 from older, traditional academic seeds. Because the simulation’s item universe is the retained corpus, it characterizes internal reachability and redundancy among retrieved sources; it does not estimate coverage of literature that the search did not retrieve. Affinity parameters represent modelled discovery opportunity rather than measured citation probabilities, and results should be read as a structural sensitivity probe rather than as a recall estimate. A.13.2 Simulation Results. The simulation results, visualized in Figure 1, are consistent with the structural stability of the chosen seed set and characterize the reachability structure of the seed network: • Rapid Saturation (Panel A): The cumulative fraction of retained items discovered per snowball iteration saturates quickly. The mean discovery rate reaches 95% expected coverage by the third snowball iteration, consistent with the depth of the manual search. • LOSO Robustness (Panel B): To test vulnerability to initial seed selection, we calculated the fraction of items still reachable when each seed is systematically removed. Overall corpus coverage remains highly robust, staying above 95% under any single seed removal. However,

56

Afdideh, Seoane, and Abtahi

Fig. 1. Snowball sampling saturation and seed-sensitivity simulation. MC analysis (5,000 replicates) evaluating the structural robustness of the 18-seed literature network. (A) Cumulative fraction of retained items discovered per snowball iteration (shading indicates 95% confidence intervals). Mean discovery reaches 95% saturation by iteration 3. (B) Leave One Seed Out (LOSO) coverage. Bars indicate the fraction of total items still reachable when a specific seed is removed. Orange markers isolate coverage strictly to the removed seed’s own chain, highlighting concentrated dependencies (e.g., RAG category retention drops to 40% without Lewis et al.). (C) Expected coverage as a function of seed-set size, demonstrating diminishing returns beyond approximately 15 seeds. (D) Mean per-item redundancy (number of unique seeds capable of reaching each item) categorized by taxonomy chain. The dashed line denotes the single-seed-dependence threshold (<1.5), which isolated categories like RAG fall below. (Note: Reachability is modeled on category-level affinity probabilities representing discovery opportunity, rather than strict historical citation traces).

isolating the coverage to the removed seed’s own chain reveals concentrated dependencies. For example, removing Lewis et al. drops RAG-specific coverage to approximately 40%, demonstrating why certain foundational papers were strictly mandatory inclusions. • Diminishing Returns (Panel C): Expected coverage plotted against seed-set size demonstrates severe diminishing returns beyond approximately 15 random seeds. The purposive selection of 18 seeds safely captures the asymptote of this discovery curve. • Network Redundancy (Panel D): Mean per-item redundancy tracks how many unique seeds can eventually reach a given item. While most taxonomy categories enjoy robust, multi-seed discovery paths, structurally isolated categories (such as RAG) fall below the 1.5

A Six-Dimensional Taxonomy of Post-Training Adaptation Techniques with Applications in AI Governance

57

single-seed-dependence threshold, further validating the necessity of a highly purposive, multi-chain seed strategy over a single, centralized query. B

Detailed Dimensional Category Definitions

This section provides the granular definitions for the categories of our proposed taxonomy. By detailing the D1 (Mechanism), D2 (Goal), D3 (Data Requirements), D4 (Persistence), and D5 (Scope), these tables allow for a clearer, multi-dimensional view of the current adaptation landscape. Table 23. D1 (Mechanism of Adaptation) Categories Category

Detailed Definition I. Gradient-Based / Compute-Intensive

Parametric Update

Cross-Model Transfer

A gradient-based optimization procedure that computes loss-driven parameter gradients—using a task-specific or dataset-specific objective—and applies them to alter the stored weight tensors of one or more model components. The mechanism can operate across the full parameter space simultaneously (full FT), or be constrained to a designated subset of layers, adapter matrices, or task-specific heads (partial and parameter-efficient variants); in all cases it consumes forward and backward passes—and therefore training-time floating-point operations—to modify the model’s learned representations. No architectural constraints are assumed: the defining property is that GD acts directly on stored parameters. A training procedure in which a new, independently deployable student model is optimized to replicate the output distributions, intermediate representations, or decision boundaries of a pre-existing teacher model. Rather than updating the teacher’s parameters, the mechanism generates a derived artifact—the student—by minimizing a surrogate objective (e.g., KL divergence over soft logits, feature-map matching, or contrastive alignment) that transfers the teacher’s generalization surface without requiring access to the original training data. The student and teacher are structurally distinct deployable artifacts. II. Gradient-Free / Weight Manipulation

Parameter Composition Parametric Compression

Architectural Modification

A gradient-free algebraic operation applied directly to the parameter vectors of one or more pre-trained model checkpoints—such as task-vector addition, negation, or weight averaging—that derives a new behavioral configuration by combining or transforming weight-space representations without executing any training loop on task data. No forward passes over labeled examples and no loss function are required; the mechanism operates entirely in parameter space. A structural or numerical intervention that reduces a single existing model’s parameter count, memory footprint, or arithmetic precision—through techniques such as magnitude pruning, structured sparsification, or weight quantization—producing a deployment-optimized derivative artifact. Unlike cross-model transfer, no independent teacher model is required; the compression operation modifies the model’s representational capacity directly by eliminating or reducing its own parameter tensors. A topological alteration of the model’s computational graph that introduces new nodes, routing pathways, or expert sub-networks—such as dynamically initializing MoE routing modules, inserting gating components, or appending new expert layers—without modifying the pre-trained weights of existing base components. The mechanism expands the model’s capacity or specialization range through structural graph surgery rather than parameter optimization on existing tensors. III. Inference-Time / Zero-Footprint

Context Injection

ActivationSpace Manipulation

InferenceTime Search

An inference-time conditioning mechanism that prepends, appends, or structurally assembles task-relevant information—including natural-language instructions, retrieved documents, few-shot demonstrations, or structured tool outputs—into the model’s input context window without modifying any stored parameter weights. Behavioral changes are achieved entirely through the model’s existing attention and generation mechanisms acting on the enriched input; no gradient computation occurs, and the model’s parameter state is identical before and after the inference session. A mechanism that adds learned steering vectors or direction-specifying perturbations to the intermediate activation representations of a model during the forward pass—typically targeting specific residual stream positions or attention head outputs—thereby redirecting the model’s behavioral trajectory along a target direction without modifying any stored weight tensors. Because the intervention operates in activation space rather than parameter space, the model’s checkpoint is entirely unaffected; the effect is confined to the forward pass and leaves no trace in the persistent parameter state. An inference-time compute-allocation mechanism that scales response quality by executing multiple internal generation processes without modifying model parameters, injecting external retrieval context, or steering activations. The mechanism trades additional inference compute for improved output quality by exploring a broader portion of the model’s generative distribution and selecting the highest-quality candidate according to a process-internal scoring criterion. Compute is consumed at inference time rather than training time. IV. Pipeline-Mediated

PipelineMediated

An indirect adaptation pathway in which the technique modifies the composition, ordering, labeling, or synthesis of training examples supplied to a downstream learning algorithm without directly modifying any model parameter. Because the technique operates on the data pipeline rather than on the model artifact itself, the persistence and structural scope of the resulting adaptation are entirely determined by and inherited from the downstream training method it feeds.

58

Afdideh, Seoane, and Abtahi

Table 24. D2 (Goal of Adaptation) Categories Category

Definition I. Capability

Build Capability Task Specialization Capability Extension

Reasoning

The objective of constructing a model with generalizable predictive or generative capability from a large, representative training corpus, establishing the baseline parameter state and intended-use scope from which all post-training adaptation techniques depart. This goal is realized during initial training and serves as the foundational reference point against which all subsequent adaptation events are characterized. The objective of refining a general-purpose or pre-trained model to achieve high performance on a specific downstream task by optimizing on task-labeled data within an existing distribution, trading breadth of generalization for depth of domain-specific accuracy. The resulting model exhibits a narrowed operational envelope that is more reliable within the target task but may show reduced generalization to substantially out-of-distribution inputs. The objective of broadening a model’s operational scope by adding fundamentally new structural capabilities or modalities—such as new languages, reasoning domains, subject areas, or input/output modalities—through additional training on broad corpora or through architectural expansion. Distinguished from task specialization by the breadth rather than depth of the intended behavioral change: the model acquires genuinely new operational dimensions rather than deepening performance within an existing one. The objective of enhancing a model’s capacity for structured multi-step inference—including mathematical problemsolving, logical deduction, code generation, and systematic planning—by training against verifiable programmatic reward signals that assess correctness independently of human preference annotation. The mechanism typically involves RL against outcome-based oracles such as unit-test executors, mathematical verifiers, or formal logic checkers, enabling the model to develop generalizable problem-solving strategies rather than pattern-matching memorized solution templates. Verification is accomplished through programmatic oracles, simulators, or iterative self-competition rather than human raters. II. Lifecycle

Drift Remediation

Continual Adaptation

Distributional Gap Bridging ∗

The objective of restoring or sustaining a deployed model’s predictive accuracy or output consistency when the operational data distribution diverges from the training distribution through covariate shift, concept drift, or prompt distribution change. Remediation targets preservation of the model’s original task objective within its established operational envelope rather than reshaping that envelope for a new task; the behavioral goal is to re-align the model with its original specification, not to extend it. The objective of enabling a model to sequentially accumulate knowledge from new tasks or data streams without catastrophic forgetting of previously consolidated representations, by employing stability-plasticity trade-off mechanisms that constrain updates to preserve prior knowledge while accommodating new distributional information. The commitment to accumulation rather than replacement is the defining property distinguishing this goal from drift remediation. The objective of reducing the feature-space mismatch between a source domain—on which the model was originally trained—and a target domain—on which it must perform—for the same or an analogous task class. The adaptation leverages unlabeled or lightly labeled target-domain data to align feature representations or adjust decision boundaries without requiring complete retraining from scratch. III. Trust

Alignment

Safety

Robustness Explainability Fairness Reliability

The objective of steering model outputs to conform with human values, behavioral preferences, helpfulness norms, and institutional guidelines, typically by optimizing against human preference signals or rule-based reward models that encode normative expectations about appropriate responses. Alignment interventions modify the model’s behavioral disposition and output style rather than its factual knowledge base or core task capability. The objective of reducing the production of harmful, dangerous, or policy-violating outputs by training the model to recognize and refuse harmful requests, applying constitutional-principle enforcement, or filtering outputs through harm-category constraints. Safety-targeted interventions explicitly shape the model’s refusal and harm-avoidance behavior as a primary optimization target, distinct from general alignment with preferences. The objective of hardening a model against adversarial perturbations, distributional noise, or out-of-distribution inputs by exposing it to worst-case or adversarially constructed training examples. The adaptation tightens the model’s minimum guaranteed performance across the input space, reducing the gap between average-case and worst-case behavior. The objective of adapting a model or appending interpretive modules to expose internal representations, align reasoning traces with human-interpretable logic, or generate post-hoc rationales. The objective of reducing systematic performance or output disparities across demographic or otherwise protected groups by intervening on the model’s learned representations or decision boundaries. The objective of aligning the model’s expressed confidence scores with empirically observed outcome frequencies, ensuring that predictions assigned a given probability are correct at correspondingly consistent rates. Reliability interventions typically apply post-hoc calibration procedures to the model’s raw output logits, without modifying its internal learned representations. IV. Knowledge Operations

Knowledge Update Remove Knowledge

The objective of injecting, refreshing, or correcting factual or domain-specific knowledge encoded in the model’s parameters or made accessible at inference time, either by retrieving current information from an external corpus at inference time or by performing targeted parametric edits that overwrite specific outdated or incorrect factual associations within the model’s weights. The objective of selectively erasing specific learned associations, factual claims, behavioral patterns, or capability traces from a trained model without requiring full retraining and without degrading unrelated model behaviors. V. Operational Constraints

Computational Efficiency Personalization Privacy Preservation Behavior Control

The objective of reducing a model’s deployment cost, memory footprint, inference latency, or training compute requirements while preserving task accuracy within an acceptable degradation tolerance. Efficiency interventions alter the model’s deployment profile without changing its core task definition or behavioral envelope. The objective of tailoring model outputs, tone, or domain expertise to individual user preferences, histories, or contexts through discrete per-user parameters or tracking mechanisms without requiring broad model-wide retraining that would affect all users uniformly. The objective of enabling collaborative model training or adaptation across distributed data holders while guaranteeing that sensitive raw data never leaves its originating environment. This is achieved by sharing only privacy-bounded gradient updates or model deltas rather than underlying data records, enabling a global model to benefit from distributed learning signals without centralizing sensitive information. The objective of constraining a model’s output format, stylistic register, persona, or procedural behavior through input-context conditioning without targeting the model’s underlying knowledge representation or task capability. Behavioral control operates at the inference interface and produces no persistent parametric modification.

A Six-Dimensional Taxonomy of Post-Training Adaptation Techniques with Applications in AI Governance

Table 25. D3 (Data Requirement) Categories Category

Definition I. Human-Annotated Data

Large beled

La-

Small Labeled

Few Demonstrations Preference Pairs Task Distribution

A high-volume corpus of annotated input-output pairs providing the statistical coverage necessary to update all model parameters reliably without overfitting. This data regime is the defining requirement for methods where the unrestricted optimization landscape requires broad supervision to constrain the high-dimensional parameter search space. A limited collection of annotated task-specific instances sufficient for parameter-efficient adapter training or partial FT but insufficient for full-parameter optimization without catastrophic forgetting or overfitting. The small scale constrains the effective update surface and necessitates parameter-efficient or regularized training strategies that limit the degrees of freedom available to the optimizer. A minimal set of input-output demonstration examples provided directly within the model’s inference-time context window to condition its behavior without any gradient-based optimization. These examples function as behavioral exemplars that the model’s in-context learning mechanism generalizes from within the active session, leaving the model’s parameters entirely unmodified. A dataset of paired model outputs annotated by human raters to indicate relative behavioral preference according to quality, helpfulness, or safety criteria. Unlike task-labeled data that specifies a correct output, preference pairs encode comparative judgments between two candidate responses, making the comparative elicitation methodology—rather than absolute correctness—the defining characteristic of this data regime. A distribution over multiple heterogeneous tasks to expose the model to diverse adaptation episodes. The distributional coverage over task space—rather than depth of supervision on any single task—is the defining property; the model optimizes for initial parameters or learning procedures that generalize rapidly to novel tasks encountered at deployment. II. Synthetic / Programmatic Data

TeacherDerived ConstitutionDerived EnvironmentDerived InstructionResponse Corpus

Data generated by a more capable or aligned model acting as an annotator or oracle (teacher). The generating model’s biases, capability limitations, and output distribution are inherited by the student through this data pathway, making provenance characterization of the teacher a prerequisite for characterizing the data. Data generated iteratively based on an explicit set of textual guidelines or constitutional principles, in which a model critiques and revises its own outputs according to the specified normative rules. The constitutional text itself functions as the primary normative artifact encoding the behavioral constraints that propagate into the adapted model through the generated training data. Data generated through interaction with a physics engine, game simulator, formal environment, or through the model’s own autoregressive self-generation acting as an opponent or world model. The simulation parameters, reward mechanics, and environmental transition dynamics define the statistical properties of the generated data; the learning signal is derived from outcome-based environmental feedback rather than human judgment. A heterogeneous collection of prompt-response pairs spanning many tasks and domains used to train general instructionfollowing capability. The defining characteristic is corpus breadth across task types rather than depth within any single task; III. Live / Production Data

User Interaction Data Test-Only Sequential / Incremental ∗

Logged behavioral signals and implicit telemetry generated directly by end-users during active deployment used to drive personalization or implicit feedback-based adaptation. This regime is operationally distinct from curated labeled datasets and comparative preference pairs in that it is generated passively during production use rather than collected through an explicit annotation pipeline. The live, unlabeled test instances encountered during active inference, used as the sole input signal for transient self-supervised adaptation without access to any external training data or ground-truth labels. The model adapts to each test batch or individual instance in a session-local manner; adaptation does not persist beyond inference completion. A temporally ordered stream of training examples arriving in batches or continuously over the model’s operational lifecycle, whose distributional properties evolve across time. Adaptation mechanisms operating on this data regime must integrate new information without discarding previously consolidated knowledge; IV. External / Targeted Data

External Corpus Forget Set Specification

A curated external knowledge repository queried at inference time to retrieve contextually relevant passages that augment model outputs. The corpus is a pre-existing, structurally static reference artifact that is independent of the active user session and of the model’s internal parameter state. A precisely delineated set of training samples, factual associations, behavioral capability traces, or user data records whose influence is to be erased from model behavior. V. Minimal / Decentralized / Delegated Data

Zero-Shot

ParameterOnly Decentralized PipelineDependent Unlabeled †

Natural-language task directives provided in the model’s input context without any accompanying labeled examples or demonstrations, relying entirely on the model’s pre-existing instruction-following and generalization capability. The data requirement is zero labeled instances: only the task specification itself is provided, and all behavioral conditioning is achieved through the model’s previously learned representations. A data-free regime in which adaptation operates exclusively on the model’s pre-trained parameter vectors through algebraic composition, interpolation, or negation—requiring no training examples, labeled or unlabeled, as input to the adaptation procedure. The entire adaptation is derived from arithmetic relationships among existing weight tensors. Private data distributed across multiple non-collocated clients or institutions, each processed locally without sharing raw records with a central coordinator. The statistical heterogeneity across shards—arising from non-independent and non-identically distributed local distributions—constitutes the primary technical challenge for any aggregation mechanism that attempts to synthesize a coherent global model from local signals. A delegation regime in which the effective data requirements of the technique are fully determined by and inherited from the downstream training method that the technique augments. The pipeline technique itself imposes no independent data regime; all data characterization is subsumed by the consuming training procedure. Unannotated data used for self-supervised or unsupervised training objectives, continuous pre-training, or distribution sampling for post-training calibration or activation steering. The absence of ground-truth labels defines this regime: the learning signal is derived from the statistical structure of the raw data—through masked prediction, contrastive objectives, or density estimation—rather than from explicit human annotation. VI. Verifier-Defined Ground Truth

Verifiable Task

Tasks for which correctness can be assessed by a deterministic programmatic verifier without requiring human preference annotation. Ground truth is established through binary or structured programmatic evaluation rather than subjective human judgment, shifting the epistemological basis of the training signal from comparative human preference to objective computational verification.

Paired Multimodal

Aligned cross-modal paired samples used to train cross-modal alignment or grounding objectives. Each training example carries independent provenance chains for each modality, and the cross-modal correspondence itself introduces additional alignment-specific data characteristics.

VII. Multi-Modal Paired Provenance

59

60

Afdideh, Seoane, and Abtahi

Table 26. D4 (Persistence) Categories Category

Definition I. Pre-Specifiable Baselines

Scheduled Permanent

Pre-planned, periodic parameter updates executed on a fixed temporal or data-volume cadence, resulting in a permanently updated model checkpoint that fully supersedes the prior parameter state. The modification is permanent at the model-parameter level—the updated weights replace the previous checkpoint and persist across all subsequent inference sessions—and its triggering conditions and approximate timing can be declared in advance.

Bounded Cumulative

Incremental parameter accumulation across a finite, pre-enumerable set of new tasks or domains, in which each update extends the model’s knowledge without overwriting prior representations. The finite, pre-specified character of the task sequence means the total knowledge accumulation horizon can, in principle, be declared in advance, making this the only cumulative persistence mode whose modification trajectory admits complete pre-specification.

VersionPersistent

A context-only modification at the model-parameter level—such as a maintained retrieval corpus, a persistent steering-vector store, or a versioned production system prompt—that does not alter model weights but is maintained as a version-controlled, persistent system artifact whose update constitutes a system-level change event. While the model’s internal parameter state is unchanged, the system’s behavioral profile is durably modified by the maintained artifact across inference sessions. II. Event-Driven Unplanned Modification

Ad-hoc Permanent

An unplanned, event-driven parameter update triggered by sudden data drift, an emergent vulnerability, a targeted knowledge correction, or an arbitrary developer decision rather than a pre-specified schedule. Like scheduled permanent updates, the change produces a new model checkpoint that permanently supersedes the prior parameter state; unlike them, the timing, trigger, and scope of the intervention cannot be declared in advance and are determined reactively. III. Continuous Surveillance Required

Unbounded Cumulative

Open-ended, continuous parameter accumulation over an indefinite or unknown horizon of streaming data, in which the model’s behavioral state is never stable but continuously incorporates new representational commitments. The absence of a pre-specifiable termination boundary is the defining property: unlike bounded cumulative persistence, the total scope of adaptation cannot be declared in advance, and the model’s behavioral state at any future checkpoint is structurally indeterminate at deployment time. IV. Session-Scoped

SessionEphemeral

A behavioral modification achieved entirely through input-context conditioning or internal generation processes at inference time, with the model’s parameter state remaining identical before and after each session. The adaptation is local to the active inference request or session and clears completely upon session termination, leaving no trace in the model’s stored parameter state. While ephemeral at the parameter level, the system-level behavioral effect may be functionally persistent if the conditioning mechanism is applied uniformly across all deployed inference sessions.

Transient

A modification strictly confined to a temporary inference session that does not persist once that session concludes. The transient category encompasses two mechanistically distinct implementations: parametermodifying variants (e.g., test-time adaptation), in which gradient updates are applied during inference and then explicitly rolled back upon session completion via weight restoration; and activation-modifying variants (e.g., activation steering), in which no weight write occurs and the effect naturally clears at session end. V. Inherited Burden

PipelineDependent

A persistence profile in which the temporal durability of the adaptation is not intrinsic to the technique itself but is entirely delegated to and inherited from the downstream learning method that the technique augments. Because the technique operates on the training data pipeline rather than directly on model parameters, its effective change-control classification is entirely contingent on the downstream training method it feeds; no independent persistence characterization applies to the pipeline technique in isolation.

A Six-Dimensional Taxonomy of Post-Training Adaptation Techniques with Applications in AI Governance

Table 27. D5 (Scope) Categories Category

Definition I. Full-System Validation

Model

Whole-

Fused Composition

All layers and components of the model are modified simultaneously, producing the most comprehensive structural footprint in the taxonomy. Because the entire parameter space participates in the modification, no modular isolation boundary exists; any behavioral change introduced propagates across all task dimensions simultaneously, and the model’s complete representational state is overwritten or jointly updated. Rollback requires full-checkpoint reversion rather than targeted component substitution. A derived model artifact produced by the algebraic combination of two or more parent model checkpoints, resulting in a merged weight state in which the individual contributing models are no longer trivially separable. The fused artifact’s behavior is not directly attributable to any single parent model, making its representational lineage structurally opaque relative to the source checkpoints. II. Targeted Component Validation

Partial Surrogate Model Compressed Artifact

Only a designated component or layer subset of the model is modified while the remainder of the parameter space is frozen. The bounded modification surface restricts behavioral change to the updated component subset, enabling targeted validation and localized regression testing of only the modified layers; unmodified layers require no re-evaluation. A completely new, independently deployable model artifact trained to replicate a teacher model’s outputs or to provide a learned scoring signal as an intermediate training artifact. Because it is a structurally new model produced as a byproduct of the adaptation procedure, it constitutes a distinct artifact from the adapted model itself and carries an independent representational provenance. A structurally reduced, pruned, or lower-precision derivative of the original base model. that occupies a strictly smaller representational capacity than the original while targeting preservation of the original model’s functional behavior within an acceptable accuracy tolerance. The compressed artifact is evaluated as a modified derivative of the original model rather than an independently trained entity. III. Modular Isolation

Modular

The adaptation is encapsulated as a decoupled, independently stored module that can be attached to or detached from a frozen base model without altering the base model’s parameter state in any way. The base model’s existing representational state is completely preserved, and the modular component constitutes an independent artifact whose versioning is entirely distinct from the base checkpoint. Rollback is achieved by removing or swapping the module without any base-model reversion. IV. No Parametric Footprint

Input/OutputSpace ActivationSpace

No modification is made to any model parameter or internal activation state; behavioral conditioning occurs exclusively through external input-space manipulation, output candidate selection, or inference-time search over the model’s generative distribution. The model’s computational graph and all stored weight tensors remain completely unaltered. The adaptation footprint is confined to the interface between the model and its deployment environment, not to any internal model state. Direct modulation of the model’s intermediate forward-pass representations with zero modification to any underlying parameter weight tensors. While the model’s internal activation state is transiently altered during the forward pass, no persistent change to any stored weight tensor occurs; the model’s checkpoint is identical before and after the intervention. V. Distributed Aggregation

Distributed

The global model state is constituted by the aggregation of locally computed parameter updates from multiple decentralized clients, such that the adaptation signal is distributed across a federated system. The persistent artifact is the aggregated global model whose parameter state incorporates learning signals from all contributing local data shards without any raw data leaving the originating device. Validation focuses on the consistency of the aggregated global state rather than on any single local update. VI. Inherited Burden

PipelineDependent

The structural scope of the adaptation is not intrinsic to the technique itself but is entirely delegated to and inherited from the downstream training event that the technique feeds. Because the pipeline technique modifies the data distribution rather than the model’s computational graph, it imposes no independent structural footprint; the component-level regression testing burden is entirely subsumed by the downstream training method that consumes the modified data.

61

62

C C.1

Afdideh, Seoane, and Abtahi

Technique Definitions and Relationships Knowledge Transfer and Task Specialization (Rows 2–7 in Table 8)

TL as Umbrella Paradigm. TL is not a single technique but the paradigm governing the reuse of knowledge acquired on a source task or domain to improve performance in a new target context [153, 234]. It functions as the conceptual container for FT, DA, and FSL (Rows 2–7), which are its principal realizations. Because its D1 (Mechanism) and D4 (Persistence) coordinates inherently vary by implementation, TL is excluded from the centerpiece table and treated as an umbrella genus whose species are individually profiled. The hierarchy PEFT ⊂ FT ⊂ TL encodes the key taxonomic relationship: each level of specialization tightens the constraints on D5 (Scope) and D3 (Data Requirements) simultaneously, and it is precisely these two dimensions that govern structural validation scope and data-provenance obligations in system-engineering documentation. Full FT (Row 2) is the maximally expressive parametric realization of TL: GD ( d1 = Parametric Update ) applies to every model layer without exception, producing a d5 = Whole-Model footprint that carries the highest structural validation burden in the taxonomy. The unrestricted update surface structurally necessitates the d3 = Large Labeled data regime—with all parameters free to drift, the optimization landscape has sufficient degrees of freedom to overfit or catastrophically corrupt prior knowledge unless constrained by large-scale supervision [108]. The d2 = Task Specialization goal draws the sharpest coordinate boundary with Retraining (Row 8): both are mechanistically identical across D1 (Mechanism), D4 (Persistence), and D5 (Scope), yet retraining targets d2 = Drift Remediation on a fixed behavioral envelope rather than reshaping that envelope for a new task. Compared to Partial FT (Row 3), Full FT sacrifices the reduced testing surface (D5 (Scope) contracts from d5 = Whole-Model to d5 = Partial ) in exchange for broader update expressiveness; compared to PEFT (Row 4), it eliminates the d5 = Modular isolation boundary that permits component-level rollback, replacing it with whole-model entanglement. Because Full FT modifies the entire parameter space with no modular isolation boundary, rollback requires complete checkpoint restoration rather than targeted component substitution, and the validation obligation extends to full-system behavioral regression testing across all previously certified behavioral boundaries before any re-deployment. Partial FT (Row 3) targets d2 = Task Specialization via the same d1 = Parametric Update mechanism as Full FT (Row 2) but constrains updates to later, task-specific layers while freezing the early layers that encode broadly transferable representations [222]. This architectural discipline collapses the update surface from d5 = Whole-Model to d5 = Partial —a coordinate divergence from Full FT that carries both a data consequence and a validation consequence: the data requirement contracts to d3 = Small Labeled (the frozen layers perform feature extraction, leaving only the task-specific layers requiring supervision), and the regression testing footprint is structurally isolated to the modified layer subset, enabling targeted component-level rollback. Compared to PEFT (Row 4), which introduces a fully decoupled d5 = Modular adapter that can be removed without touching any base-model layer, Partial FT embeds updates directly into the base model’s later layers—rollback requires layer-level checkpoint restoration, not a clean module swap. The practical distinction matters for change-control documentation: a PEFT event generates a standalone adapter artifact; a Partial FT event generates a modified base-model checkpoint whose changed layers must be explicitly versioned. Partial FT necessitates versioned checkpointing of the modified layer subset and localized parametric regression testing bounded strictly to those layers; because the frozen early layers remain mathematically unchanged, they require no re-validation. PEFT (Row 4) is taxonomically the most distinctive member of the FT family, differentiated by its d5 = {Partial, Modular} scope, which no full or partial FT variant achieves. The defining operational property is that the base model’s weights remain strictly frozen throughout training; all adaptation occurs within a small set of newly introduced parameters appended to or inserted into the frozen

A Six-Dimensional Taxonomy of Post-Training Adaptation Techniques with Applications in AI Governance

63

architecture [90]. This produces the d5 = Modular coordinate: the adapter is a structurally independent artifact that can be loaded, swapped, or removed without touching the base checkpoint. This d5 = Modular property creates a structural bridge to Task Arithmetic and Model Merging (Row 32), where adapter task-vectors can be algebraically combined. The data requirement contracts to d3 = Small Labeled relative to Full FT because the adapter’s parameter count is orders of magnitude smaller than the base. At the LLM and MLLM tiers, PEFT simultaneously absorbs the distributional gap goal of DA (Row 6) when applied to domain-specific corpora, and its modular architecture makes it the preferred complement to alignment techniques (RLHF, DPO) in production stacks. Because a change-control event in a PEFT deployment affects only the adapter artifact—not the frozen base—regression testing scope is bounded to the adapter module alone; the base model’s existing validation record remains unmodified, enabling clean rollback by removing the adapter checkpoint without any base-model reversion. While categorized as a single structural decision unit in the taxonomy, the PEFT paradigm encompasses a wide array of architectural variants. These include quantization-aware methods (e.g., QLoRA [52]), continuous prompt interventions (prefix-tuning, prompt-tuning, p-tuning), sparse bias updates (BIas-Term FIne-Tuning (BitFit) [25]), and attention-scaling (IA3 [125]). Furthermore, dynamic adapter composition strategies, such as adapter fusion and LoRA routing (e.g., X-LoRA [35], LoraHub [94], S-LoRA [177]), functionally blur the boundary between modular PEFT and MoE (Row 34) by dynamically routing inputs to specialized adapter modules. ReFT (Row 5) presents an alternative to standard weight-based PEFT by intervening on the internal activation streams of the model [216]. While it utilizes a d1 = Parametric Update , it does not alter the base weights; instead, it trains lightweight, d5 = Modular projection matrices (interventions) that alter the model’s hidden representations during the forward pass. It targets d2 = Task Specialization using d3 = Small Labeled datasets. Because it introduces permanent, swappable components, it maintains d4 = {Scheduled Permanent, Ad-hoc Permanent} persistence. It is distinct from Activation Steering (which calculates transient vectors) because ReFT permanently trains its intervention matrices via GD. DA (Row 6) addresses d2 = Distributional Gap Bridging —reducing the feature-space mismatch between source and target domains when labeled target data is unavailable or scarce. Its d3 = {Unlabeled, Small Labeled} profile reflects the defining operational condition: the target domain lacks sufficient labeled examples to support task-specific FT, The d1 = Parametric Update mechanism operates by training a feature extractor alongside an explicit domain discriminator or statistical divergence minimizer, yielding a d5 = Partial footprint (the alignment layers and feature extractor are updated; task heads may remain frozen) [66]. The defining taxonomic boundary is d6 = {DL, FM} : at the LLM tier, the DA goal persists but the mechanism is superseded—CPT (Row 27) achieves distributional alignment by continued self-supervised training on unlabeled target-domain corpora, and domain-specific PEFT (Row 4) injects target-domain signal via lightweight adapters. System documentation for LLM-level DA must therefore specify the concrete CPT or PEFT pathway rather than invoking classical DA terminology. At the d6 = {DL, FM} tier, change-control documentation should specify which alignment layers were updated and which domain-discriminator architecture was used, as these define the precise structural boundary of the modification event; regression testing should verify that source-domain performance is preserved alongside target-domain improvement—necessitating a dual-distribution evaluation protocol as part of the validation record. FSL (Row 7) is defined within this taxonomy exclusively as a d1 = Parametric Update intervention under extreme data scarcity. Depending on the implementation, it relies on either 1–5 examples for metric-based support sets ( d3 = Few Demonstrations ) or tens of examples for gradient-based updates ( d3 = Small Labeled ), producing a persistent weight modification ( d4 = {Scheduled Permanent, Ad-hoc Permanent} ) [185, 201]. This strict boundary resolves the taxonomy’s most consequential terminological ambiguity: colloquial “few-shot” usage conflates two mechanistically and operationally distinct operations. The

64

Afdideh, Seoane, and Abtahi

parametric form produces a versioned model artifact that has undergone a parameter-modifying training event and therefore demands formal change-control tracking and baseline regression testing. The non-parametric, prompting-based form shares d3 = Few Demonstrations but operates via d1 = Context Injection , producing only a d4 = Session-Ephemeral behavioral shift that leaves the underlying model artifact unmodified; this is reassigned to ICL (Row 39). The D4 (Persistence) divergence is the taxonomic anchor: FSL creates a new versioned artifact requiring software change-control; ICL creates no artifact whatsoever. FSL’s d6 = {DL, FM} ceiling mirrors DA’s (Row 6): at the LLM tier, ICL achieves rapid few-shot task specialization through context conditioning alone, making formal FSL training pipelines architecturally redundant for pure few-shot tasks at that scale. FSL necessitates version-control tracking of the resulting weight checkpoint; the extremely small d3 = Few Demonstrations dataset constitutes the sole training provenance record and should be fully documented. C.2

Temporal Adaptation and Maintenance (Rows 8–11 in Table 8)

Retraining (Row 8) updates ( d1 = Parametric Update ) a deployed model on newer or additional data to restore performance degraded by distribution shift, representing the foundational lifecycle maintenance operation across all model types. Its universality across d6 = {ML, DL, FM, LLM, MLLM} reflects reflects the fact that distributional drift is an architecture-agnostic failure mode: any model whose training distribution diverges from its deployment distribution will degrade, regardless of architectural family [65]. The critical D2 (Goal) divergence from (Full/Partial) FT (Rows 2–3) is precise: both operate identically on D1 (Mechanism) and D4 (Persistence), yet retraining preserves the model’s original task objective ( d2 = Drift Remediation ) rather than reshaping its capability boundaries for a new task ( d2 = Task Specialization ). This D2 (Goal) distinction carries a direct validation consequence—a retraining event that strictly adheres to the original task specification does not require an intended-use review; an FT event does. The d3 = Sequential / Incremental coordinate reflects that retraining data arrives in temporally ordered batches whose distributional properties evolve across the model’s operational lifecycle—this is structurally distinct from the static, curated datasets of (Full/Partial) FT ( d3 = {Small Labeled, Large Labeled} ) or FSL (Row 7; d3 = {Small Labeled, Few Demonstrations} ). The d4 = {Scheduled Permanent, Ad-hoc Permanent} coordinate marks the boundary with CL (Row 9): unconstrained retraining overwrites prior weight states with a new fixed checkpoint, whereas CL’s d4 = {Bounded Cumulative, Unbounded Cumulative} encodes a commitment to accumulation rather than replacement. The d5 = Whole-Model scope means that every weight is overwritten at each retraining event; there is no modular isolation boundary, so any behavioral drift introduced by the new training data propagates across all task dimensions simultaneously—necessitating automated full-system regression testing against a frozen behavioral baseline before re-deployment, with the historical checkpoint retained as the rollback artifact. CL (Row 9) addresses the stability-plasticity trade-off [108, 136]: how to accumulate knowledge across sequential tasks or shifting distributions without overwriting previously learned representations. Its defining coordinate is D4 (Persistence): d4 = {Bounded Cumulative, Unbounded Cumulative} —a deliberate architectural commitment to knowledge accumulation rather than replacement—which is the main axis that separates CL from Retraining (Row 8, d4 = {Scheduled Permanent, Ad-hoc Permanent} ) despite their identical D1 (Mechanism) and D3 (Data Requirements) profiles. The bounded variant corresponds to sequential acquisition over a finite, pre-enumerable task set (TIL, Row 10); the unbounded variant corresponds to open-ended streaming adaptation over an indefinite sequence of distributions (DIL, Row 11). This cumulative persistence coordinate is operationally consequential: the open-ended modification trajectory of d4 = Unbounded Cumulative fundamentally breaks the assumptions of discrete, periodic regression testing, because the model’s behavioral state is never stable—it is continuously accumulating new representational commitments. At the LLM tier, CL

A Six-Dimensional Taxonomy of Post-Training Adaptation Techniques with Applications in AI Governance

65

encompasses three sequential stages—continued pre-training, domain-adaptive pre-training, and continual FT [178]—each carrying distinct D3 (Data Requirements) and D5 (Scope) profiles. The D5 (Scope) coordinate ( d5 = Partial ) reflects that production CL deployments at scale almost always apply constrained, parameter-isolated updates (adapters, selective layer freezing) rather than wholemodel retraining, precisely to limit the forgetting surface. CL’s cumulative persistence necessitates continuous behavioral drift monitoring rather than periodic snapshot regression testing; each task or distribution boundary must be logged as a discrete lifecycle event with its own targeted stability evaluation verifying that prior-task performance has not regressed beyond a predefined tolerance. TIL (Row 10) is a specific operational variant of CL in which a model acquires a finite, preenumerable sequence of distinct tasks—each presented without access to prior task data—with task identity explicitly provided at inference time [199]. The d4 = Bounded Cumulative coordinate is the main definitional anchor: the finite, pre-specified task sequence means the total knowledge accumulation horizon can, in principle, be declared in advance, making TIL the only CL variant whose change trajectory is pre-specifiable in a technical plan. Its d5 = {Partial, Modular} scope reflects that the dominant practical implementations at the DL and FM tiers use parameter-isolated architectures—dedicated output heads, task-specific adapters, or masked parameter subsets—that physically separate each task’s learned representation from prior ones, directly mitigating the forgetting problem through structural disentanglement rather than regularization. The d6 = {DL, FM} ceiling is imposed by architectural supersession: at the LLM tier, domain-adaptive pre-training and continual PEFT subsume TIL’s goals under the broader CL (Row 9) framework, rendering dedicated TIL protocols redundant. Compared to DIL (Row 10), TIL’s inference-time task identifier is a system-level dependency that must be tracked in deployment documentation—if the task-identity signal is unavailable or corrupted, the deployed system’s behavior is undefined, constituting a unique failure mode that requires explicit specification in the operational validation plan. TIL’s bounded, pre-specifiable task sequence supports formal pre-deployment documentation of the complete capability accumulation trajectory; the modular scope permits per-task regression testing at each task boundary, isolating validation to the newly added task module without re-validating prior task components. DIL (Row 11) is the open-ended ( d4 = Unbounded Cumulative ) variant of CL: the model encounters a potentially indefinite sequence of domains or class distributions without access to prior training data and—critically—without a task identifier at inference time [199]. This absence of inference-time task identity is the defining coordinate divergence from TIL (Row 10). The d5 = {Partial, Whole-Model} scope reflects that some DIL approaches retrain the entire network while others selectively update domainadaptive components. The compound D2 (Goal) profile— d2 = {Continual Adaptation, Drift Remediation} —captures DIL’s dual operational goal: it simultaneously accumulates new distributional knowledge ( d2 = Continual Adaptation ) and preserves performance on prior distributions that may re-emerge ( d2 = Drift Remediation ). The d6 = {DL, FM} ceiling is imposed by the same supersession mechanism as TIL: at the LLM tier, continued pre-training and domain-adaptive PEFT subsume DIL’s goals. DIL’s unbounded modification trajectory is the most operationally demanding persistence profile in the taxonomy: the system’s behavioral state cannot be fully characterized at any fixed checkpoint, necessitating continuous runtime monitoring with per-distribution performance tracking and explicit drift-impact assessments at each domain boundary transition. C.3 Alignment, Reasoning, and Trustworthiness (Rows 12–19 in Table 8) SFT (Row 12), also termed instruction tuning, establishes the foundational behavioral capability for an LLM to interact via natural language instructions [152], and its taxonomic position demands precise separation from classical task-specific FT (Rows 2–4). Both employ d1 = Parametric Update via supervised GD, but the D3 (Data Requirements) divergence is categorical: task FT operates on

66

Afdideh, Seoane, and Abtahi

d3 = Small Labeled datasets of narrowly scoped, single-function examples yielding a bounded operational envelope, whereas SFT employs a massively heterogeneous d3 = Instruction-Response Corpus —spanning question answering, summarization, code generation, reasoning chains, and multilingual instructions—to instill a general instruction-following capability that cuts across all task boundaries. This D3 (Data Requirements) divergence propagates into the D2 (Goal) profile: SFT achieves a multi-domain form of d2 = Task Specialization alongside baseline d2 = {Alignment, Safety} , rather than the single-function specialization of task FT. While traditional SFT relies on human-curated datasets, contemporary variants frequently utilize synthetic, model-generated data [209] (e.g., Instruction Backtranslation [120], Reward rAnked FineTuning (RAFT) [54], and Self-Rewarding models [225]). This introduces a critical D3 (Data Requirements) nuance: it shifts the compliance and bias-auditing burden away from human annotator governance and onto the upstream teacher model generating the synthetic corpus. Safety SFT is a specific instantiation that explicitly enforces d2 = Safety by restricting the corpus to refusal demonstrations and harm-category filtered examples. The D5 (Scope) coordinate— d5 = {Partial, Whole-Model} —is implementation-dependent: SFT applied to all parameters yields d5 = Whole-Model with a full-system regression burden. Because SFT reshapes the model’s behavioral envelope across all tasks simultaneously—not just a single bounded function—it demands comprehensive cross-domain behavioral regression testing before re-deployment; a task-specific regression suite is insufficient; the validation protocol must cover the full instruction-following surface, including all task types represented in the instruction-response corpus. RLHF (Row 13). RLHF trains a reward model on human preference pairs and optimizes an LLM policy against it, constituting the canonical alignment technique for generative models [43]. The d3 = Preference Pairs coordinate is the mechanism’s defining input requirement and its most consequential data-governance obligation: unlike d3 = Small Labeled datasets annotated for a fixed objective, preference pairs encode comparative behavioral judgments—which of two model outputs a human annotator prefers—making annotator demographics, selection criteria, and inter-rater reliability first-class artifacts in the training provenance record. The d2 = {Alignment, Safety} goal and the strict d6 = {LLM, MLLM} bound reflect that preference-based alignment is architecturally contingent on the model’s capacity for open-ended natural-language generation; the preference signal is structurally meaningless for discriminative or shallow architectures. The D5 (Scope) coordinate is implementation-dependent in a way that carries direct validation consequences: applying RL policy optimization across all network parameters yields d5 = Whole-Model , demanding full-system regression testing; applying it via PEFT adapters restricts the scope to d5 = Partial , isolating the change-control footprint to the adapter; this structural distinction is operationally critical, as it determines whether an alignment modification event demands full-system regression testing or a targeted, component-level validation assessment. Training the reward model introduces a third d5 = Surrogate Model component—a fully independent model artifact that triggers its own conformity validation chain. This last point is the taxonomically critical divergence from DPO (Row 14): both share D1 (Mechanism), D2 (Goal), D3 (Data Requirements), and D4 (Persistence), but RLHF’s reward model produces a d5 = Surrogate Model footprint that DPO eliminates entirely by recasting alignment as a direct classification objective. RLHF deployments require independent validation of the reward model artifact as a separate system component; any update to the reward model constitutes a distinct change-control event from the policy update, as the reward model defines the optimization target for all subsequent policy training. DPO (Row 14) eliminates the explicit reward model required by RLHF, recasting alignment as a supervised classification objective directly on preference pairs [164]. DPO shares D1 (Mechanism), D2 (Goal), D3 (Data Requirements), and D4 (Persistence) with RLHF (Row 13)—making the D5 (Scope) divergence the sole but operationally decisive difference between them. DPO’s elimination of the reward model collapses D5 (Scope) to d5 = {Partial, Whole-Model} alone, removing the d5 = Surrogate Model

A Six-Dimensional Taxonomy of Post-Training Adaptation Techniques with Applications in AI Governance

67

component entirely. This D5 (Scope) reduction is the most system-engineering-consequential property of DPO: it eliminates the need for a separate reward model validation pipeline, and the three-stage orchestration infrastructure. Compared to RLAIF/CAI (Row 15), DPO shares the same D5 (Scope) simplification but retains the d3 = Preference Pairs human-annotation requirement; RLAIF substitutes AI-generated constitutional preferences ( d3 = Constitution-Derived ), shifting the validation burden from annotator oversight to constitutional text auditing. DPO’s structural simplicity narrows the validation surface to a single policy artifact; however, the preference pair dataset remains the sole provenance record determining the alignment boundary, and its annotator-selection protocol, demographic coverage, and inter-rater reliability metrics must be fully documented as the system’s behavioral specification. Recent advancements have expanded this family into broader preferenceoptimization variants, including IPO [22], KTO [58], and SimPO [142]. Notably, ORPO [86] collapses the traditional multi-stage pipeline by integrating baseline instruction tuning and alignment into a single gradient update. RLAIF/CAI (Row 15) share the optimization infrastructure of RLHF but fundamentally substitutes the D3 (Data Requirements) profile [111]: human-annotated preference pairs are replaced by synthetic preferences generated by an AI model operating under a set of constitutional principles ( d3 = Constitution-Derived ) [23]. This D3 (Data Requirements) substitution is the technique’s defining taxonomic property and its consequential system-engineering decision: it transfers the entire source of the alignment signal from human annotators—whose demographics, biases, and consistency are tracked through inter-rater reliability metrics—to a text-based constitutional document and an AI feedback model whose behavioral properties, latent biases, and edge-case failure modes must instead be audited. Compared to RLHF (Row 13) and DPO (Row 14), RLAIF/CAI shares D1 (Mechanism), D2 (Goal), D4 (Persistence), D5 (Scope), and D6 (Model Type) but diverges exclusively on D3 (Data Requirements)—making the data-source substitution the main structural change. The scalability advantage is architectural: AI annotation parallelizes without human bottlenecks, enabling preference dataset generation at orders-of-magnitude larger scale than human annotation pipelines permit. Because the constitutional text and the AI feedback model jointly constitute the alignment specification, both must be treated as version-controlled engineering artifacts; any modification to either triggers a new change-control event, and the validation protocol must include semantic-stability testing of the constitutional principles across edge-case behavioral scenarios and regression testing of the AI critic pipeline’s preference consistency. RLVR/GRPO (Row 16) targets d2 = Reasoning —multi-step problem-solving, planning, and chainof-thought verification—a goal that diverges fundamentally from both the alignment goal of RLHF/DPO/RLAIF ( d2 = {Alignment, Safety} ) and the d2 = Task Specialization goal of FT [74, 175]. The D3 (Data Requirements) coordinate is also a differentiator: d3 = Verifiable Task relies on programmatic reward signals (e.g., mathematical correctness verified by symbolic solvers, code execution verified by test suite pass/fail, logical entailment verified by formal proof checkers) rather than subjective human preference annotation. This substitution does not merely simplify the data pipeline; it categorically changes the nature of the validation obligation. The preference-pair regime (RLHF/DPO) requires mitigating annotator bias and tracking demographic representation; the verifiable-task regime requires instead certifying the correctness and coverage of the programmatic oracle—a shift from human-factors verification to formal functional-specification testing. The D3 (Data Requirements)-based differentiation between RLVR and RLHF is the clearest demonstration of the multi-dimensional taxonomy’s analytical value: a mechanism-only single-axis framework collapses both into a generic “RL-based tuning” category, completely obscuring that they target distinct D2 (Goal) goals and require radically different D3 (Data Requirements) provenance documentation. LP (Row 17) is framed as an interpretability and explainability intervention—determining its D2 (Goal) assignment as d2 = Explainability —rather than merely as a downstream evaluation metric, task

68

Afdideh, Seoane, and Abtahi

performance, or behavioral alignment [222]. By freezing all base model parameters and training only a lightweight linear classifier appended to a designated intermediate layer, LP surfaces latent representational structure as interpretable task-specific outputs without modifying the base model’s core feature extraction function [16]. The d1 = Parametric Update classification applies exclusively to the probe head’s weight matrix; the base remains mathematically unchanged. Its d5 = Modular scope indicates that LP probes constitute structurally decoupled artifacts whose installation and removal leave the base model’s parametric state and validated behavior unaffected. Compared to PEFT (Row 4), which also introduces a modular artifact onto a frozen base, LP diverges critically on D2 (Goal): PEFT targets task performance ( d2 = {Task Specialization, Computational Efficiency} ), while LP targets representational transparency. Compared to LEACE (Row 18), which operates via d1 = Parameter Composition to algebraically rewrite the base model’s weight matrices, LP never touches the base weights at all. For production deployments requiring verifiable transparency, the probe’s classification performance across sensitive concept dimensions can be documented as engineering evidence of representational alignment or bias. Because LP does not modify the base model’s parameters, it does not trigger a model-level change-control event; the decoupled probe artifact ( d5 = Modular ) requires its own version-controlled documentation specifying the target layer, concept labels, and labeled probe dataset—but the base model’s existing validation record remains entirely unaffected. LEACE (Row 18) targets d2 = Fairness through a mechanistically novel D1 (Mechanism) pathway: rather than employing iterative GD, it operates via d1 = Parameter Composition —computing a covariancebased orthogonal projection from a d3 = Small Labeled and algebraically projecting the model’s weight matrices into the null space of a target concept’s linear representation [24], a parameter composition operation analogous to Task Arithmetic (Row 32) but operating on concept subspaces rather than task capability directions. This closed-form linear algebra operation bypasses the computational footprint of traditional parameter-updating methods entirely, producing a d4 = Ad-hoc Permanent , d5 = Partial modification at negligible training cost. Compared to LP (Row 17), which trains a probe head on top of a frozen base to expose linear concept encodings, LEACE inverts the operation: it takes the same linear concept subspace identified by probing and algebraically projects it out of the base model’s weight matrices—making LP and LEACE complementary, with LP as the diagnostic step and LEACE as the corrective intervention targeting the same identified representation. The d4 = Ad-hoc Permanent Permanent coordinate reflects that LEACE produces an irreversible, event-driven modification at negligible compute cost—unlike gradient-based fairness interventions that require training runs, LEACE can be applied as a post-hoc weight surgery. The d5 = Partial scope is bounded by the targeted layers: only the weight matrices encoding the concept’s linear subspace are projected; all other parameters remain unchanged. Its operational and validation significance lies in the deterministic auditability of the intervention: because the modification is derived analytically from a strict concept specification, it is fully reproducible and mathematically characterizable; the concept subspace, projection matrix, and labeled probe dataset constitute a complete audit trail for the fairness intervention—validation requires verifying that the modified representation is linearly indecipherable for the target concept while baseline task performance is preserved, both of which are statistically testable properties. Adversarial training (Row 19) improves model robustness by incorporating adversarially perturbed examples directly into the optimization objective through a min-max formulation [134]. The d1 = Parametric Update coordinate reflects the defining operational property: adversarial perturbations are generated dynamically within the optimization loop, meaning the gradient signal itself is adversarially shaped at each training step—this is the mechanistic distinction from Data Augmentation (Row 25), which modifies training inputs as a static pre-processing step independent of the gradient loop. This distinction is taxonomically consequential: adversarial training is a gradient-coupled technique whose computational cost scales with the inner maximization procedure at each step,

A Six-Dimensional Taxonomy of Post-Training Adaptation Techniques with Applications in AI Governance

69

while Data Augmentation’s cost is absorbed in the data-loading stage. Both share d2 = Robustness , but only adversarial training carries d1 = Parametric Update with an adversarially shaped gradient signal, rather than the d1 = Pipeline-Mediated mechanism of augmentation. The d6 = {DL, FM, LLM} ceiling reflects an active research frontier: cross-modal adversarial robustness for vision-language MLLM architectures lacks standardized post-training protocols and established benchmark coverage, making the MLLM tier currently out of scope. Adversarial training universally trades off standard operational performance against robustness [195]; this Pareto trade-off must be explicitly characterized in the validation record, with the robustness-accuracy frontier documented as a system performance boundary—any deployment threshold revision following adversarial training constitutes a system behavior change requiring separate regression validation. C.4

Training Strategies (Rows 20– 24 in Table 8)

Meta-learning (Row 20) frames adaptation itself as an optimization objective: rather than learning to perform a single task, the model learns an initialization, metric space, or learning procedure that enables rapid generalization to new tasks from minimal data ( d3 = Task Distribution ) [63, 87]. Its D3 (Data Requirements) coordinate is the defining characteristic: instead of a labeled dataset for one task, meta-learning requires a diverse distribution of training tasks—a data-provenance structure qualitatively different from the task-specific corpora of Full FT ( d3 = Large Labeled ) or PEFT ( d3 = Small Labeled ). Its D6 (Model Type) ceiling at d6 = {DL, FM} encodes a significant structural observation within the taxonomy: at the LLM tier, ICL (Row 39) has architecturally supplanted gradient-based meta-learning as the primary mechanism for rapid few-shot adaptation. The supersession is mechanistically grounded: LLMs contain sufficiently rich pre-trained representations that few-shot task generalization can occur through in-context conditioning ( d1 = Context Injection , d4 = Session-Ephemeral ) rather than through expensive parametric optimization ( d1 = Parametric Update , d4 = {Scheduled Permanent, Ad-hoc Permanent} ). This supersession relationship dictates that system engineers deploying LLMs should route rapid adaptation requirements through prompt-level ICL architectures, reserving formal meta-learning pipelines exclusively for DL and early FM deployments. The D4 (Persistence) coordinate is counterintuitive for a technique framed around adaptability: metalearning produces a permanent parametric modification ( d4 = {Scheduled Permanent, Ad-hoc Permanent} )—not a transient or session-ephemeral one. This means meta-learning demands the same rigorous full-system regression testing and baseline validation protocols as any standard training event, structurally distinct from ICL’s zero-artifact execution. Meta-learning requires that the task distribution used for outer-loop training be fully documented as the primary provenance record; the diversity and coverage of this task distribution directly determines the generalization bounds of the resulting meta-initialization, and any modification to the task distribution constitutes a change event requiring re-validation of the adaptation baseline. MTL (Row 21) jointly optimizes shared representations across a fixed task set, exploiting intertask relatedness to improve generalization on each individual task [37]. The defining coordinate is d3 = Task Distribution : the training corpus spans a predetermined collection of tasks simultaneously available in memory, rather than a single-task labeled dataset (Full FT, Row 2, d3 = Large Labeled ). Its structural distinction from Meta-Learning (Row 20) is that, while both share d3 = Task Distribution , MTL’s task set is statically fixed at training time rather than dynamically drawn from a distribution. The d5 = Whole-Model scope is the critical consequence of the joint optimization: because all taskspecific output heads share the same backbone parameters, there is no modular isolation boundary between tasks—any change to the shared encoder propagates across all tasks simultaneously. This means modifying the task set, task-weighting schedule, or encoder architecture requires full-system regression testing across all tasks to verify that positive transfer to new tasks has not induced negative interference on existing ones. At the MLLM tier, MTL is the foundational

70

Afdideh, Seoane, and Abtahi

architecture underlying multimodal instruction tuning (Row 46)—a unified backbone must serve heterogeneous vision, audio, and text task heads simultaneously. MTL requires comprehensive crosstask regression testing at every task-set modification boundary; the inter-task interference structure must be explicitly characterized and documented, as negative transfer to any task constitutes a behavioral regression on that task’s validated performance envelope. Self-play (Row 22) generates training data by having a model compete against or evaluate itself, producing an auto-curriculum of increasing difficulty without requiring externally provided labeled data. The d3 = Environment-Derived coordinate is the defining structural property: the training signal is generated internally through the model’s own interactions with copies or historical checkpoints of itself—no human annotation, no static corpus, and no external oracle are involved in data generation. This self-referential data source produces the compound d2 = {Task Specialization, Alignment, Reasoning} profile—the three operationally distinct goals served by selfplay at different D6 (Model Type) levels: domain mastery through competitive co-evolution in DL game agents ( d2 = Task Specialization ) [182], instruction-following improvement through self-generated preference pairs in LLM alignment ( d2 = Alignment ) [42], and chain-of-thought reasoning cultivation through iterative self-verification in reasoning models ( d2 = Reasoning ) [226]. This multi-goal profile would be irresolvable in a single-axis taxonomy; the taxonomy’s value is precisely in making explicit that the same D1 (Mechanism) and D3 (Data Requirements) regime serve fundamentally different D2 (Goal) depending on the deployment context—and that each dictates entirely distinct evaluation metrics, regression test suites, and operational baselines. Compared to Curriculum Learning (Row 23), which also generates structured training sequences, self-play mainly differs on D1 (Mechanism) ( d1 = Parametric Update vs. d1 = Pipeline-Mediated ) and D3 (Data Requirements) ( d3 = Environment-Derived vs. d3 = Pipeline-Dependent )—Curriculum Learning sequences existing data, while self-play generates new data through model-environment interaction. The d6 = DL scope (skipping FM) reflects that standard FM FT pipelines do not typically employ self-play, which re-emerges only as a specialized adaptation mechanism for alignment and reasoning at the LLM tier. Because the specific D2 (Goal) being pursued dictates the validation protocol, any self-play deployment must explicitly document which goal dimension applies, providing a clear audit trail for the intended-purpose and risk-rationale in compliance documentation. Curriculum learning (Row 23) structures the ordering of training examples rather than their content, presenting data in an easy-to-hard sequence to improve optimizer trajectory and final convergence [26]. Its d1 = Pipeline-Mediated classification is the decisive taxonomic property: curriculum learning never executes a gradient update on the model—it operates strictly as a data-sequencing function upstream of the optimizer. The parametric modification that results from training on a curricularized dataset is performed entirely by the downstream training method; curriculum learning contributes only the temporal ordering of the input stream. This indirect mechanism is the structural reason for d4 = Pipeline-Dependent and d5 = Pipeline-Dependent : the D4 (Persistence) and D5 (Scope) profile of the resulting model are fully inherited from whichever downstream training method consumes the sequenced data. Compared to Data Augmentation (Row 25), Data Augmentation synthetically extends the training set (via transformation), while curriculum learning operates on the existing dataset without modification—it changes the sampling order, not the examples. Compared to Active Learning (Row 24), both share d3 = Pipeline-Dependent , but Active Learning introduces a d2 = Distributional Gap Bridging component by querying strategically chosen unlabeled samples, while Curriculum Learning’s d2 = Task Specialization reflects that it optimizes the convergence trajectory rather than the distributional coverage. Because curriculum learning does not independently modify model parameters, it must be documented as a named algorithmic dependency of the primary training event it serves—specifying the difficulty-scoring methodology, sequencing protocol, and any pacing

A Six-Dimensional Taxonomy of Post-Training Adaptation Techniques with Applications in AI Governance

71

functions—rather than as a standalone change-control entry; the downstream training event inherits the full parametric and structural documentation burden. Active learning (Row 24) strategically selects the most informative unlabeled samples for oracle annotation, maximizing model improvement per labeled example by targeting the data points the model is most uncertain about [174]. Like Curriculum Learning (Row 23), it carries d1 = Pipeline-Mediated : the model’s parameters are updated by a downstream training method, not by the active learning mechanism itself. However, Active Learning diverges from Curriculum Learning on D2 (Goal): where curriculum learning sequences existing labeled data for convergence optimization ( d2 = Task Specialization only), active learning’s query strategy inherently introduces a d2 = Distributional Gap Bridging component—by selectively targeting uncertainty-rich, informationally dense regions of the unlabeled distribution, it constructs a labeled dataset whose coverage is deliberately biased toward the model’s current capability boundaries. This D2 (Goal) distinction is operationally significant: curriculum learning does not change the statistical composition of the labeled dataset; active learning does, and the algorithmic selection strategy is the direct determinant of the distributional coverage of the resulting training corpus. The d4 = Pipeline-Dependent and d5 = Pipeline-Dependent coordinates confirm that persistence and structural scope are fully inherited from the downstream training method—but the labeled dataset itself, as an artifact of the active selection process, constitutes an independent data-provenance record that must be documented and audited separately. The active learning acquisition function must be explicitly documented and version-controlled as a system configuration artifact; the selection strategy directly defines the boundaries of the model’s validated operational envelope, and any change to the acquisition function constitutes a data-pipeline modification requiring re-evaluation of the distributional coverage of the resulting labeled corpus before downstream training commences. C.5

Data-Centric and Privacy-Preserving Methods (Rows 25–29 in Table 8)

Data Augmentation (Row 25) increases training data diversity through synthetic transformations of existing examples, without acquiring new labeled data [62, 181]. Its d1 = Pipeline-Mediated classification reflects its structural position: augmentation operates strictly upstream of the gradient loop, modifying training inputs before they reach the optimizer—it executes no forward pass on the model and computes no gradients. This upstream positioning means the technique shares d4 = Pipeline-Dependent and d5 = Pipeline-Dependent coordinates with all other boundary-extension methods (Curriculum Learning, Active Learning, Semi-SL), reflecting that the D4 (Persistence) and D5 (Scope) of any resulting model change are entirely determined by the downstream training method that consumes the augmented data. The compound d2 = {Task Specialization, Robustness} profile reveals a structurally important internal distinction within augmentation: standard transformations (flips, crops, synonym substitution) diversify the input distribution to improve generalization ( d2 = Task Specialization ), while adversarially constructed perturbations specifically target distributional robustness ( d2 = Robustness ) [195]. The latter variant overlaps in goal with Adversarial Training (Row 19), but differs critically on D1 (Mechanism): adversarial training generates perturbations dynamically within the gradient loop (d d1 = Parametric Update with adversarially shaped gradients), while adversarial augmentation applies fixed perturbation transforms as a static pre-processing step ( d1 = Pipeline-Mediated )—an operationally meaningful distinction because the dynamic coupling in adversarial training produces adaptive perturbations that scale with the model’s current vulnerability, while static augmentation perturbations cannot. The augmentation policy—comprising the exact set of transformation functions, their mathematical parameterizations, hyperparameter bounds, and random seed configurations—constitutes a component of the training data specification and must be version-controlled and documented as an immutable dependency of the training event; a

72

Afdideh, Seoane, and Abtahi

change to the augmentation policy redefines the statistical distribution the model optimizes against and requires a new data-provenance record. Semi-SL (Row 26) leverages abundant unlabeled data alongside a small labeled set, making it highly valuable in domains where annotation is prohibitively expensive. Its d1 = Pipeline-Mediated classification groups it with the other data-pipeline strategies, but Semi-SL occupies a structurally distinct position within this group: unlike Curriculum Learning (which sequences existing labeled data) and Data Augmentation (which transforms existing labeled data), Semi-SL introduces unlabeled data—whose labels were never verified by a human or certified oracle—as a training signal source [28, 186]. Semi-SL is uniquely complex because it integrates two distinct data sources ( d3 = {Unlabeled, Small Labeled} ). Because the unlabeled pool lacks ground-truth verification, its inherent sampling bias and distributional coverage directly shape the model’s learned behavior, creating a structural transparency gap that distinguishes Semi-SL from purely supervised pipelines. Compared to SSL/CPT (Row 27), which shares the d3 = Unlabeled dependency and uses self-supervised objectives directly, Semi-SL is d1 = Pipeline-Mediated —it does not itself prescribe an optimization objective or update parameters; it delivers a mixed labeled-and-pseudo-labeled dataset to a downstream training method. SSL/CPT is d1 = Parametric Update , defining its own objective and directly executing weight modifications. Because the unlabeled data pool’s distributional properties directly determine the scope and boundaries of the downstream model’s validated capabilities, the unlabeled corpus must be subjected to the same provenance documentation standards as the labeled set—including source identification, collection methodology, distributional characterization, and bias auditing. SSL/CPT (Row 27) occupies a structurally dual role in the taxonomy (marked † in Table 8): foundational pre-training from scratch is explicitly out of scope, but applying SSL as an adaptation technique to an already-deployed architecture is within scope. The critical coordinate that distinguishes SSL-as-adaptation from all other pipeline-mediated data strategies is d1 = Parametric Update [77]: unlike Data Augmentation (Row 25) or Semi-SL (Row 26), which modify the data pipeline and delegate parameter updates to a downstream method, CPT directly defines its own optimization objective and formally executes GD on model weights. The d3 = Unlabeled coordinate is shared with Semi-SL (Row 26), but the mechanism is structurally distinct: Semi-SL uses unlabeled data to construct pseudo-labeled training examples for a downstream training method; CPT uses unlabeled data as the direct training signal for its own self-supervised objective, making it a non-supervised parametric adaptation technique in the taxonomy. The D4 (Persistence) coordinate bifurcates based on execution mode: d4 = Scheduled Permanent when CPT is applied as a discrete offline batch on a fixed domain corpus; d4 = Unbounded Cumulative when applied as continuous domain-specific streaming—the latter creating an open-ended modification trajectory that breaks the assumptions of static regression testing and demands dynamic behavioral monitoring. At the LLM and FM tiers, the primary adaptation strategy leveraging SSL objectives is CPT: applying self-supervised objectives (e.g., nexttoken prediction) to unlabeled, domain-specific corpora to internalize novel vocabulary, syntax, and factual structures without requiring supervised labels. CPT functions as an initialization phase preceding SFT (Row 12) or PEFT (Row 4), providing a domain-enriched representational baseline that ensures the model possesses foundational domain competency before task-specific supervision is applied [77]. DP-FT (Row 28), using Differentially Private Stochastic Gradient Descent (DP-SGD), incorporates gradient clipping and noise injection during training to provide mathematical guarantees against membership inference and data extraction attacks. DP-FT is a d1 = Parametric Update dedicated exclusively to the d2 = Privacy Preservation goal. Unlike FL, which achieves privacy via d5 = Distributed data sovereignty, DP-FT operates on centralized data ( d3 = {Small Labeled, Large Labeled} ) but mathematically protects the records. Like standard parametric updates, DP-FT produces a permanently modified weight checkpoint that can be deployed on a planned cadence or as an event-driven patch, assigning

A Six-Dimensional Taxonomy of Post-Training Adaptation Techniques with Applications in AI Governance

73

it d4 = {Scheduled Permanent, Ad-hoc Permanent} . It can be applied to the d5 = Whole-Model or as Differentially Private Parameter-Efficient Fine-Tuning (DP-PEFT) ( d5 = Partial ). Because the DP-SGD algorithm is foundational, its applicability spans deeply into the hierarchy, yielding d6 = {DL, FM, LLM, MLLM} . FL (Row 29) trains a global model by aggregating gradient updates from local models on decentralized datasets, ensuring raw training data never leaves local devices [103, 137]. Its d1 = Parametric Update classification reflects that FL does execute GD and produce formal weight modifications—distinguishing it from pipeline-mediated data strategies—but the critical structural differentiator is d3 = Decentralized : no centralized training corpus exists. Each client computes gradients locally on its private data shard and transmits only model updates (weight deltas or gradients) to the server for aggregation; raw data never leaves the originating device. This decentralized data geometry is what justifies d2 = Privacy Preservation as a D2 (Goal): FL does not merely apply privacy as a constraint layered onto a task-training objective—privacy preservation is an intrinsic structural property of the data access pattern, not an add-on mechanism. The absence of centralized training data means that data-quality auditing, distributional characterization, and bias assessment must be delegated to client-local verification processes or statistical reporting protocols; the server cannot directly inspect the training corpus, making third-party data-provenance attestation mechanisms a system-engineering requirement rather than an optional audit supplement. C.6

Efficiency and Composition (Rows 30–34 in Table 8)

KD (Row 30) transfers a trained teacher model’s learned behavioral distribution into a smaller student model, most commonly by training the student to match the teacher’s soft probability output distribution, enabling compression of capability without commensurate compression of performance [72, 84]. Its d1 = Cross-Model Transfer classification captures the defining structural property: a complete, separately trained model is the proximate information source for the student’s training—not a labeled dataset, not a teacher’s parameters directly, but the teacher’s output distribution. This cross-model information flow distinguishes KD from all other compression-adjacent techniques: Model Compression (Row 31) operates on a single model’s own parameters while KD requires a functioning teacher model to generate training signal, making d3 = Teacher-Derived . The d2 = Computational Efficiency goal drives the student’s architectural reduction (smaller, faster inference). From a validation standpoint, a KD event establishes a strict behavioral dependency chain: verifying the student model requires comprehensive technical documentation of both the teacher model’s original training data provenance and the exact algorithmic distillation protocol used to generate the soft targets, ensuring full auditability of the behavioral alignment downstream. The student artifact inherits the teacher’s behavioral envelope but introduces a compression-induced approximation gap; this gap must be characterized quantitatively—via behavioral regression testing across the teacher’s validated operational envelope. Model compression (Row 31) reduces the computational footprint of an already-trained model through structural modification of its parameters, producing a deployable compressed artifact with lower memory and inference cost [81]. Its d1 = Parametric Compression classification explicitly encodes reduction as the defining operation: unlike all d1 = Parametric Update techniques that modify parameters to improve capability, d1 = Parametric Compression modifies parameters to reduce resource consumption while attempting to preserve capability. Its d2 = Computational Efficiency coordinate designates an objective of footprint reduction that inherently lacks a target for intrinsic capability development. Compared to KD (Row 30), which achieves efficiency through architectural reduction of a student trained from teacher outputs ( d1 = Cross-Model Transfer , d3 = Teacher-Derived ), Model Compression achieves efficiency by directly restructuring a single model’s existing parameters, i.e., no second model is required. Its D3 (Data Requirements) profile diverges fundamentally based on the chosen algorithmic implementation: Data-free compression methods operate strictly as d3 = Parameter-Only interventions;

74

Afdideh, Seoane, and Abtahi

because they consume zero training or calibration data, they completely bypass data-provenance pipelines, making them optimal for strict zero-data-access operational deployments. Conversely, calibration-dependent compression methods require a d3 = Unlabeled calibration dataset to calculate activation statistics and minimize quantization error [64]. The d5 = Compressed Artifact is a structurally altered model artifact—not a parameter overlay applied to the original—and must be registered as a distinct model version requiring its own validation suite: the compression operation introduces approximation errors whose distribution across the input space cannot be inferred from the original model’s validation record; full behavioral regression testing on the compressed artifact against the original model’s validated operational envelope is required.From a regulatory and change-control perspective, this category encompasses two dimensionally distinct approaches. Post-Training Quantization (Post-Training Quantization (PTQ), e.g., Generative Pre-trained Transformers Quantization (GPTQ) [64]) operates as a training-data-free ( d3 = Parameter-Only ) algebraic derivation applied after the model is fully trained. In contrast, Quantization-Aware Training (QAT) simulates lower precision during active GD, classifying it as a distinct d1 = Parametric Update training event with differing validation and data-provenance requirements. Task Arithmetic and Model Merging (Row 32) combine pre-trained and fine-tuned checkpoints into a single multi-capable model through arithmetic operations on parameter space, without any additional training data [100, 214, 218]. Its d1 = Parameter Composition and d3 = Parameter-Only classification suggest no gradient computation, no data consumption at merge time, no teacher model inference—only algebraic operations on weight tensors. Machine Unlearning (Row 47) shares d1 = Parameter Composition via task-vector negation (subtracting a forget-task vector to remove a learned capability), making Machine Unlearning a targeted polarity reversal of the exact same algebraic operation that Task Arithmetic uses to add capabilities. Compared to MTL (Row 21), which achieves multi-task capability through joint optimization over a shared task-distribution corpus ( d1 = Parametric Update , d3 = Task Distribution ), Task Arithmetic achieves the same compound capability profile through post-hoc parameter composition requiring no additional training and no access to the original task datasets. To mitigate parameter interference and preserve specialized capabilities during the fusion process, practitioners employ advanced mathematical merge operators, including TIES-MERGING [218], DARE [223], SLERP [131, 180], and Fisher merging [135]. Consequently, the regression validation suites used to verify targeted machine unlearning—specifically, verifying the erasure of a specific capability while ensuring the preservation of adjacent foundational knowledge—are directly applicable to Task Arithmetic (task-vector operations). Because Model Merging bypasses traditional training telemetry, system validation must rely on downstream functional testing and post-merge automated regression benchmarks to verify that parameter-space interference has not degraded core model capabilities. Long-Context Extension (Row 33) methods (e.g., Position Interpolation (PI) [40], Yet another RoPE extensioN (YaRN) [156], and Periodic Extrapolation Positional Encodings (PEPE) [91]) extend a language model’s operational context window by adapting its positional encoding (e.g., through Rotary Positional Embedding (RoPE) [189] scaling) and, when required, fine-tuning on long-sequence data, enabling the model to process inputs far exceeding its original context length. This operates via a composite D1 (Mechanism): it requires an d1 = Architectural Modification to alter the frequency scaling of the RoPE, typically followed by a brief d1 = Parametric Update on long documents. The goal is strictly d2 = Capability Extension , as it structural expands the model’s capacity rather than teaching a new task. It targets the d5 = Partial scope and utilizes d3 = {Unlabeled, Small Labeled} long-sequence corpora. Because it permanently alters the positional embedding architecture and corresponding parameters, it produces a versioned checkpoint exhibiting d4 = {Scheduled Permanent, Ad-hoc Permanent} persistence. Unlike techniques that require generation mechanics, context scaling can be applied to base GenFMs, establishing its D6 (Model Type) footprint at d6 = {FM, LLM, MLLM} .

A Six-Dimensional Taxonomy of Post-Training Adaptation Techniques with Applications in AI Governance

75

MoE Adaptation (Row 34) introduces sparse gating over a population of expert sub-networks or integrating novel expert sub-networks into an existing MoE, enabling a model to route each input to the subset of parameters most relevant to that input’s characteristics—achieving scalable multi-task specialization without proportionally scaling inference cost [56, 157, 176]. This structural distinction necessitates a compound D1 (Mechanism) profile: the topological addition of a new expert constitutes d1 = Architectural Modification —physically restructuring the computational graph—while the subsequent tuning of expert weights or the central routing mechanism constitutes d1 = Parametric Update . This graph-level intervention is what mandates d5 = Modular : the model is a collection of individually addressable expert modules coordinated by a router—each expert constituting an independently auditable behavioral unit whose activation patterns can be tracked per input. Compared to PEFT (Row 4), which adds lightweight parameter overlays (adapters, LoRA) to a frozen base with d5 = {Partial, Modular} , MoE Adaptation adds structural routing logic that fundamentally changes the model’s computational topology—adapter modules are additive overlays; expert modules are gated substitutes for dense computation. The compound D2 (Goal) profile— d2 = {Capability Extension, Computational Efficiency} —reflects a unique operational property: MoE adaptation is the only technique in the taxonomy that simultaneously scales systemic capacity while explicitly preserving fixed inference latency bounds. The d1 = Architectural Modification classification has stringent system-engineering implications in the taxonomy: a new model topology is being registered, not a parametric override of an existing one—requiring a complete re-specification of the model’s computational graph and expert routing logic audit trail as components of the system configuration record, since any change to router architecture or gating threshold alters which computation paths are activated for which inputs. C.7

Inference-Time Adaptation (Rows 35–42 in Table 8)

PE (Row 35) designs input text to guide a deployed model’s output distribution entirely through natural language instructions placed in the context window, without requiring any parameter modification, representing the lowest-overhead adaptation method in the taxonomy [34, 126]. Its d1 = Context Injection classification suggests that it produces no persistent model artifact whatsoever—no checkpoint, no parameter delta, no weight version, no trained embedding. The behavioral effect exists only while the prompt occupies the active context window ( d4 = Session-Ephemeral ) and operates exclusively at the model’s input/output interface ( d5 = I/O Space ). Compared to ICL (Row 39), which shares d1 = Context Injection and d4 = Session-Ephemeral but adds d3 = Few Demonstrations to steer taskspecific output patterns through example-based conditioning, PE operates with d3 = Zero-Shot through instruction specification alone—relying on generalization from the model’s pre-training rather than in-context exemplar matching. This zero-data, zero-gradient profile positions PE as the d2 = Behavior Control technique in the taxonomy that imposes the lowest modification overhead but provides the weakest behavioral guarantee: steering is non-parametric and entirely contingent on the model’s pre-existing instruction-following capability. Compared to Prompt Learning (Row 36), which replaces discrete natural-language tokens with continuously differentiable embedding vectors optimized via GD ( d1 = Parametric Update , d3 = Small Labeled ), PE operates entirely in human-legible natural language—making it interpretable but non-optimizable. The d6 = {LLM, MLLM} coordinate reflects that effective PE requires a model with emergent instruction-following capability; below LLM scale, models lack the generalization from instruction composition. Because PE produces no auditable architectural or parametric artifact, every deployed system whose behavior is governed by a prompt template must specify that template as a versioned configuration artifact; any change to the prompt—including rewording, structural reordering, or instruction addition—must be treated as a system configuration change requiring re-evaluation of behavioral coverage and output boundary characterization. The system-level variant—where core system prompts or metaprompts

76

Afdideh, Seoane, and Abtahi

are hardcoded into production deployment infrastructure—exhibits d4 = Version-Persistent persistence. This boundary requires rigorous tracking within software configuration management and release engineering protocols, as changes to the global system prompt dictate the functional behavioral envelope of the entire application despite the underlying static model weights. Prompt Learning (Row 36) converts the discrete, human-authored natural-language prompt into a set of continuously differentiable embedding vectors—soft tokens—prepended to the model’s input representation and optimized via GD on a task dataset [113, 121]. It resembles PE in its input-space operation—prepending sequence tokens to the input—but functions mechanically like FT in its optimization: learnable continuous embeddings (soft tokens) are explicitly trained via backpropagation ( d1 = Parametric Update ).Therefore, the parametric modification is confined entirely to the prepended embedding vectors ( d5 = Partial ); all base model parameters remain frozen. The d4 = {Scheduled Permanent, Ad-hoc Permanent} coordinate means trained soft prompt tensors must be registered as distinct parametric artifacts in the system configuration record: deploying a different soft prompt tensor with the same frozen base model constitutes a parametric modification to the deployed system and requires version control and behavioral regression testing against the task coverage specification defined during original validation. APO (Row 37) frameworks, such as Declarative Self-improving Python (DSPy) [106] or Optimization by PROmpting (OPRO) [219], replace manual prompt engineering with algorithmic search, treating language models as programmable engines. APO utilizes d1 = Context Injection within the d5 = Input/Output-Space , meaning the underlying model parameters remain completely frozen. However, because the optimization compiles a tested, metric-optimized prompt pipeline, the resulting artifact is classified as d4 = Version-Persistent rather than d4 = Session-Ephemeral . Data requirements are flexible: APO can utilize d3 = Few Demonstrations to bootstrap the optimization, but it can also operate strictly d3 = Zero-Shot by relying entirely on programmatic evaluation metrics or an "LLM-as-ajudge" [115, 231] to score unannotated inputs. Ultimately, this algorithmic search aims for both d2 = Task Specialization (optimizing accuracy on a target task) and d2 = Behavior Control (systematically enforcing strict output formats, stylistic guidelines, or reasoning templates without altering weights). RAG (Row 38) augments LLM generation by retrieving relevant documents from an external knowledge base and injecting them into the input context at inference time [114]. Its nonparametric nature ( d1 = Context Injection ) and d4 = Version-Persistent coordinate confirm that RAG strictly does not modify underlying model parameters: the injected knowledge alters the model’s transient output without touching its foundational weights. This fundamentally distinguishes RAG from Knowledge Editing (Row 48), which permanently rewrites factual associations within the weight space itself. The information source is a separately maintained, indexable knowledge repository whose content is decoupled from the model’s weights and can be updated independently ( d3 = External Corpus )—documents can be added, revised, or removed from the corpus. This decoupling is the sharpest structural contrast with CPT (Row 27), which internalizes domain knowledge directly into model weights ( d1 = Parametric Update , d4 = {Unbounded Cumulative, Scheduled Permanent} ): CPT makes knowledge intrinsic and persistent at the cost of a full retraining cycle for every knowledge update; RAG makes knowledge external and updateable at inference time at the cost of retrieval latency and context-window occupation. The compound d2 = {Task Specialization, Knowledge Update} profile reflects the dual function: the retrieval mechanism enables the model to answer with temporally current information ( d2 = Knowledge Update ) while the retrieved context shapes the generation to the specific domain and task at hand ( d2 = Task Specialization ). RAG functions as a complement to FT rather than a direct substitute: FT encodes structural behaviors and stylistic patterns directly into parameters, while RAG retrieves localized factual context from an external corpus. From a system engineering and validation perspective, the retrieval corpus constitutes a critical external data dependency. Consequently, the quality, currency, and provenance of the vector database must be rigorously

A Six-Dimensional Taxonomy of Post-Training Adaptation Techniques with Applications in AI Governance

77

version-controlled; updating the retrieval corpus constitutes a significant system-level configuration change that dictates the model’s factual operational envelope and demands end-to-end integration testing, even when the core model parameters remain entirely unmodified. ICL (Row 39) leverages a deployed LLM’s pattern-matching capability to perform task adaptation from a small number of input-output demonstrations placed in the context window, with no gradient computation and no parameter modification [34]. It shares a common structural mechanism—D1 (Mechanism), D4 (Persistence), D5 (Scope), D6 (Model Type)—with PE, but fundamentally differs on D3 (Data Requirements)—the inclusion of d3 = Few Demonstrations —which shifts its primary D2 (Goal) focus toward d2 = Task Specialization . From a system validation perspective, because ICL leaves the foundational model parameters entirely unchanged, it triggers no artifact-level regression testing. Instead, the specific demonstration sets act as transient input dependencies and must be documented. CE (Row 40) is the broader architectural discipline of dynamically assembling an LLM’s inference context from multiple retrieved, historical, or real-time data sources—encompassing external document retrieval [67], agent memory [154], user profiles, tool outputs, and conversational history [154] ( d3 = {External Corpus, User Interaction Data} ). CE differs from PE in the structural locus of design: PE optimizes the content of a static input prompt, while CE designs the automated retrieval pipelines and orchestration systems that dynamically populate the model’s context window at inference time. CE’s d3 = {External Corpus, User Interaction Data} coordinate strictly distinguishes it from PE ( d3 = Zero-Shot ) by explicitly demanding complex, version-controlled retrieval and state-management infrastructure. Crucially, the live inference query serves solely as an execution trigger and data constraint, never as an on-the-fly parametric gradient signal ( d3 ≠ Test-Only )—a strict architectural boundary that fundamentally separates CE from TTA (Row 41). From a system engineering perspective, CE shifts the validation burden away from artifact-level model testing and toward complex state-machine verification. Because the model’s operational envelope is dictated by dynamically retrieved external state, regression testing must rigorously validate the orchestration logic, context truncation algorithms, and data pipeline integrity rather than the static underlying model weights. TTA (Row 41) updates model parameters ( d1 = Parametric Update ) transiently during inference on individual test examples or execution batches ( d3 = Test-Only ), remediating distribution shift ( d2 = Drift Remediation ), systematically reverting those gradient updates before processing the next prediction ( d4 = Transient ) [202]. TTA is strictly classified at a d6 = {DL, FM} ceiling because executing transient parameter updates via backpropagation is computationally prohibitive at the LLM scale; at the LLM tier, prompt-based adaptation, e.g., PE (Row 35), ICL (Row 39), or Activation Steering (Row 49) serves the functionally equivalent role without executing parameter modification. To maintain terminological precision, TTA must be distinguished from Test-Time Training (TTT) [190]. While both adapt parameters at inference, TTA typically relies on unsupervised entropy minimization to counter distribution shift, whereas TTT executes transient gradient updates against a distinct self-supervised objective formulated directly on the test instance. From a system validation perspective, TTA’s transient modification profile technically bypasses the need for persistent, artifact-level regression testing, as the base artifact theoretically remains unaltered. However, the engineering and quality-assurance burden shifts heavily to execution-state management: the inference pipeline must mathematically guarantee and continuously verify the perfect restoration of the base model weights between inference episodes to prevent insidious cross-inference parameter leakage or progressive state corruption. Test-Time Compute Scaling (Row 42) dynamically allocates increased inference-time computation budget ( d1 = Inference-Time Search ) to systematically improve reasoning quality ( d2 = Reasoning ) without modifying any underlying model parameters ( d5 = Input/Output-Space ) [184]. Unlike TTA (Row 41), it

78

Afdideh, Seoane, and Abtahi

executes absolutely no gradient updates; unlike PE (Row 35), its performance gain derives mechanically from internal search algorithms and multi-step generation topologies ( d3 = Zero-Shot ) rather than explicit input phrasing. Its non-parametric D1 (Mechanism) classification and d4 = Session-Ephemeral coordinate confirm that it leaves the model architecture mathematically unchanged. Test-time compute scaling serves as the foundational mechanism driving modern reasoning models’ characteristic “extended thinking” behavior: by generating multiple candidate solutions, applying reward-model reranking, or algorithmically exploring a tree of intermediate reasoning steps, the system intentionally trades inference-time compute latency for output quality [221]. From a system engineering and validation perspective, test-time compute scaling is strictly an execution strategy rather than an artifact modification. Because the base model weights remain untouched, it bypasses parameter-level regression testing. However, because dynamic search algorithms fundamentally alter the determinism, upper latency bounds, and operational behavioral envelope of the system, this strategy demands rigorous validation through execution-state stress testing. C.8

Calibration, Personalization, and Multimodal Adaptation (Rows 43–46 in Table 8)

Calibration (Row 43) post-hoc aligns a model’s output confidence scores with its empirical accuracy, correcting the systematic overconfidence or underconfidence produced by standard training objectives that optimize for discriminative accuracy rather than probabilistic faithfulness [73, 158]. Calibration does not improve the model’s task accuracy, it improves the correspondence between the model’s stated confidence and its actual accuracy ( d2 = Reliability ). From a system engineering and safety validation perspective, calibration operates not merely as a performance optimization, but as a critical operational safety constraint. Systems integrated into automated pipelines must mathematically bound their uncertainty; uncalibrated confidence scores inherently invalidate downstream threshold-based decision logic and fail-safe routing, mandating rigorous statistical auditing of calibration metrics. Cross-Modal Alignment (Row 44) trains a model to map representations from divergent data modalities—most commonly vision and language—into a unified, shared embedding space using contrastive optimization objectives [163]. Its d1 = Parametric Update classification applies to the alignment bridge—the projection network [124], cross-attention module [116], or contrastive alignment layer [163]—while the modality-specific encoders may be frozen. Cross-Modal Alignment requires aligned pairs ( d3 = Paired Multimodal ) from two modalities—image-caption pairs, video-transcript pairs, audio-text pairs—where the pairing reflects semantic correspondence across modalities, not merely co-occurrence. The d2 = Capability Extension coordinate reflects that Cross-Modal Alignment enables capabilities that are structurally impossible in unimodal models, capabilities that require the bridge, not merely improved unimodal encoders. Its D6 (Model Type) classification at d6 = FM reflects the immense architectural scale required: contrastive training across massive, noisy paired corpora is computationally feasible only at the FM and MLLM scale. Since the modification is concentrated in the alignment bridge ( d5 = Partial ), any change to the bridge must be treated as a capability-level modification requiring full re-characterization of the system’s multimodal behavioral envelope. Modality-Specific Adapters (Row 45) extend a frozen multi-modal model’s capabilities to new modalities or improve its performance on existing ones by inserting lightweight PEFT-style ( d1 = Parametric Update ) modules ( d5 = Modular ) within each modality’s dedicated processing pathway, without modifying the shared model backbone or cross-modal bridge [191]. In contrast to CrossModal Alignment (Row 44), which optimizes the interface between representations to enable semantic integration, Modality-Specific Adapters refine within-modality processing, making them complementary adaptation layers rather than alternatives. Compared to standard PEFT (Row 4), which applies within a single modality’s processing pathway, Modality-Specific Adapters are architecturally instantiated per modality, creating a collection of independently versioned parametric

A Six-Dimensional Taxonomy of Post-Training Adaptation Techniques with Applications in AI Governance

79

modules whose composition defines the system’s active multi-modal capability profile. Their classification strictly at d6 = MLLM reflects the architectural prerequisite: independent per-modality adapter injection inherently requires a multi-stream architecture featuring distinct, unimodal computational pathways. The d5 = Modular ensures that modality-specific adaptation artifacts can be independently versioned, tested, and rolled back. From a system engineering and validation standpoint, this structural isolation profoundly simplifies configuration management: it mathematically guarantees that high-frequency updates to one modality’s adapter (e.g., integrating a new visual encoding standard) can be executed and validated completely independent of the other (e.g., text generation), systematically bypassing the computationally expensive need for end-to-end, multi-stream regression testing. Multimodal Instruction Tuning (Row 46) fine-tunes ( d1 = Parametric Update ) a multimodal model on multimodal instruction-response pairs ( d3 = Paired Multimodal ) teaching the model to follow naturallanguage instructions [124]. Its D5 (Scope) varies by implementation ( d5 = {Partial, Whole-Model} ). The D3 (Data Requirements) coordinate of d3 = Paired Multimodal establishes a highly distinctive dataprovenance dependency: unlike text-only instruction tuning, multimodal datasets require perfectly synchronized, temporally aligned instances across modalities. This severely raises the bar for annotation quality pipelines and introduces strict cross-modal consistency as a mandatory validation criterion. From an engineering and system integration perspective, this technique constitutes a high-impact, d5 = {Partial, Whole-Model} modification event that fundamentally alters the model’s core behavioral envelope across all supported modalities simultaneously. Consequently, it mandates exhaustive, cross-modal regression testing to mathematically guarantee that optimization and capability alignment in one specific modality do not destructively interfere with or degrade baseline performance in the others. C.9

Knowledge Modification and Activation-Based Adaptation (Rows 47–49 in Table 8)

Machine Unlearning (Row 47) selectively erases learned knowledge, behavioral patterns, or training-data influence ( d2 = Remove Knowledge ) from a deployed model without retraining from scratch, addressing privacy deletion requirements, copyright compliance, and safety-critical knowledge removal ( d2 = Privacy Preservation ) [36]. Compared to LEACE (Row 18), which removes a concept’s linear influence from activations by modifying the weight matrix’s projection geometry ( d1 = Parameter-Only , applied to a concept direction), Machine Unlearning targets a broader knowledge construct—facts, training data points, behavioral patterns. Its D1 (Mechanism) coordinate spans two distinct mechanistic profiles that must be rigorously distinguished within system configuration management and regression tracking. Gradient-based approaches apply gradient ascent on the target forget set ( d3 = Forget Set Specification ) to mathematically reverse the historical learning signal, operating as d1 = Parametric Update with a d5 = Partial scope. Conversely, task-vector negation approaches algebraically subtract a localized task vector from the base model’s weights to structurally strip out a target capability, operating as d1 = Parameter Composition with d5 = Fused Composition —mechanistically identical to Task Arithmetic (Row 32) executed with a negation operator. Consequently, they must be governed by the exact same change-control validation logic and interference testing suites. Adversarial membership inference attacks, targeted behavioral probes, and jailbreak-style elicitation of the supposedly forgotten knowledge constitute some available verification tools, and none provides a formal completeness guarantee; any system deploying Machine Unlearning as a compliance mechanism must document this verification gap explicitly and specify the probe suite used to assess forgetting fidelity [93]. Knowledge Editing (Row 48) modifies a specific factual association encoded in a trained model’s parameters without retraining and ideally without disrupting the model’s knowledge of logically unrelated facts [140]. Its D1 (Mechanism) coordinate is highly implementation-dependent and

80

Afdideh, Seoane, and Abtahi

must be explicitly codified in system configuration and artifact versioning pipelines. Closed-form methods treat internal Multi-Layer Perceptron (MLP) layers as key-value stores and analytically compute rank-one weight updates, operating strictly as d1 = Parameter Composition . Conversely, metalearning and optimization-based methods apply iterative gradient updates focused on the target fact, operating as d1 = Parametric Update . Knowledge Editing structurally differs from Machine Unlearning (Row 47) primarily in operational directionality: editing corrects or injects data ( d2 = Knowledge Update ), while unlearning ablates it ( d2 = Remove Knowledge ). While both share a d4 = Ad-hoc Permanent persistence profile, their D3 (Data Requirements) differ fundamentally ( d3 = Small Labeled vs. d3 = Forget Set Specification ). Crucially, strict sequential scalability bounds remain a limiting operational boundary that must be explicitly quantified in system capability documentation [76]. Activation Steering (Row 49) controls LLM behavior by algebraically injecting computed or learned steering vectors directly into intermediate hidden activations during inference execution, strictly without modifying the underlying stored parameter weights [166, 196, 237]. A notable advancement in this space leverages SAEs to identify and manipulate interpretable latent features, providing a highly precise, semantically grounded source for generating these transient steering vectors [46]. Its d1 = Activation-Space Manipulation classification occupies a mathematically unique structural position: it is mechanistically distinct from both input-layer PE (Row 35, which operates exclusively on the discrete input token space) and traditional FT (Rows 2–4, which alters permanent weight matrices). The d4 = Transient , d5 = Activation-Space coordinate profile dictates that activation steering imposes zero permanent structural modification on the base artifact, yet enforces an immediate, highly non-linear behavioral shift during the active inference session. A profound structural insight within this framework emerges when contrasting activation steering with TTA (Row 41): while both represent transient execution-time adaptations, they physically bifurcate at the architectural scale boundary. TTA executes via transient gradient-based parameter updates at d6 = {DL, FM} , whereas activation steering operates via direct latent-space vector injection at d6 = {LLM, MLLM} . C.10

Classification Tensions

Several techniques resist clean single-category classification. These tensions are inherent to the adaptation landscape and motivate the multi-dimensional approach. FSL (Row 7). In this framework, the FSL categorization is strictly reserved for parametric approaches—such as gradient-based meta-learning or metric-based networks—where the model’s weights undergo a d4 = {Scheduled Permanent, Ad-hoc Permanent} update based on limited labeled data. While the term FSL is colloquially used to describe providing examples in an LLM prompt, this framework classifies that non-parametric, d4 = Session-Ephemeral mechanism distinctly as ICL (Row 39). ICL functions primarily to induce a d2 = {Task Specialization, Knowledge Update} mapping during inference, bypassing parameter updates entirely. DA (Row 6). Stops at d6 = {DL, FM} : the goal of bridging domain gaps persists at the LLM level, but the mechanism is superseded by CPT and domain-specific PEFT. PE vs. ICL (Rows 35 vs. 39). While structurally identical on their underlying mechanism (D1 ), scope (D5 ), and model tier (D6 ), and partially overlapping in their goals (D2 ) and persistence (D4 ), the shift in data requirements ( d3 = Zero-Shot vs. d3 = Few Demonstrations ) causes their profiles to diverge. ICL uniquely introduces d2 = Task Specialization , while PE distinctly extends to d2 = Behavior Control and d4 = Version-Persistent implementations. MoE adaptation (Row 34). Straddles architecture and adaptation; the row captures the adaptation act (training routers, adding experts) rather than the architectural pattern. RLVR vs. RLHF/DPO (Row 16 vs. Rows 13–14). All share d1 = Parametric Update but differ on D2 ( d2 = Reasoning vs. d2 = {Alignment, Safety} ) and D3 ( d3 = Verifiable Task vs. d3 = Preference Pairs ). RLVR and DPO share d5 = {Partial, Whole-Model} , while RLHF adds d5 = Surrogate Model . The boundary blurs when RLVR

A Six-Dimensional Taxonomy of Post-Training Adaptation Techniques with Applications in AI Governance

81

induces alignment-like properties or when RLHF pipelines incorporate verifiable components; in production, both are often applied sequentially to the same base model. SSL/CPT (Row 27). Has a dual role: pre-training paradigm (out of scope) and adaptation technique (in scope); unlike other data-manipulation techniques, SSL-as-adaptation prescribes its own training objective and receives d4 = {Unbounded Cumulative, Scheduled Permanent} . Machine Unlearning vs. Knowledge Editing (Rows 47 vs. 48). share D1 and D4 but differ on D2 : unlearning targets d2 = Remove Knowledge while editing targets d2 = Knowledge Update . The boundary blurs when correction effectively requires erasing the old association. Activation Steering vs. TTA (Rows 49 vs. 41). Both produce d4 = Transient adaptations but differ on model type ( d6 = {LLM, MLLM} vs. d6 = {DL, FM} ), suggesting transient adaptation bifurcates at the FM/LLM boundary. Machine Unlearning and Task Arithmetic (Rows 47 vs. 32). Share d1 = Parameter Composition (task-vector manipulation): unlearning via negation is structurally identical to Task Arithmetic with a minus sign, raising the question of whether unlearning is a distinct technique or a specific application of Task Arithmetic. LP (Row 17). A strictly frozen linear probe does not modify the base model’s internal representations or predictive behavior; it merely surfaces intermediate states. It is included in the adaptation framework strictly because appending a probe fundamentally alters the deployed system’s explainability footprint and output characteristics ( d2 = Explainability ), triggering regulatory transparency mandates (e.g., EU AI Act Article 13) even when the core parameters remain untouched. Knowledge editing (Row 48). Straddles two fundamentally different structural mechanisms depending on the algorithm deployed. Closed-form editors (ROME [140], MEMIT [141]) algebraically rewrite weight matrices ( d1 = Parameter Composition ), whereas gradient-based editors (MEND [145]) rely on compute-intensive backpropagation ( d1 = Parametric Update ). D

Exploratory Structural Consistency Analysis

To rigorously verify the structural consistency of the characterizations defined in Table 8, we performed distance-based clustering analysis to quantify the internal diversity and topological distribution of our taxonomy. D.1

Methodology

We calculated a pairwise Gower-style dissimilarity matrix for all 48 techniques across D1 –D6 . 1. Jaccard Distance: The dimension values were parsed into native mathematical sets. The distance was calculated using the Jaccard distance to faithfully preserve set-subset relationships (e.g., preventing techniques applicable to d6 = LLM and d6 = {LLM, MLLM} from being treated as strictly orthogonal): |𝐴 ∩ 𝐵| 𝑠 6 (𝑖, 𝑗) = 1 − |𝐴 ∪ 𝐵| 2. Final Unweighted Gower Distance: The fundamental formulation of Gower distance allows for variable weighting across dimensions. In this analysis, we applied equal weighting (𝑤𝑘 = 1) across all six dimensions. The overall pairwise distance 𝐷 (𝑖, 𝑗) between two techniques was computed as the arithmetic mean of their six independent dimensional distances: Í𝐾 𝑤𝑘 𝑠𝑘 (𝑖, 𝑗) 𝐷 (𝑖, 𝑗) = 𝑘=1 , Í𝐾 𝑘=1 𝑤 𝑘 where 𝑖, 𝑗 ∈ {1, . . . , 48} and 𝐾 = 6. This mean operation serves a critical normalizing function. By averaging the dimensional penalties, we ensure that the global distance strictly bounds to [0, 1],

82

Afdideh, Seoane, and Abtahi

and more importantly, we guarantee that no single dimension dominates the topological projection. A divergence in a technique’s Mechanism (𝐷 1 ) exerts the exact same geometric repulsive force in the latent space as a divergence in its Goal (𝐷 2 ) or Persistence (𝐷 4 ), ensuring the UMAP projection faithfully reflects the holistic multi-dimensional profile. We subsequently projected the Gower distance matrix into a 2D latent space using UMAP (using a fixed random seed, n_neighbors=7, and metric=’precomputed’) to preserve both local neighborhoods and broader topological gradients. D.2

Topological Overlap of Functional Families

Fig. 2. Structural consistency analysis of the adaptation taxonomy. (Left) Pairwise Gower distance heatmap of the 48 techniques, grouped by functional family. Light blocks indicate low Gower distance (high structural similarity); dark blocks indicate high distance (structural orthogonality). (Right) 2D UMAP projection colored by the 9 predefined functional families.

The Gower heatmap (Figure 2; left) provides a raw mathematical view of the structural overlaps. The presence of lighter-colored blocks across different family boundaries confirms that these categories are not isolated silos but share significant mechanical traits. This high-dimensional proximity is geometrically realized in the UMAP projection (Figure 2; right). This visual distribution confirms that the functional families defined in Section C occupy complex, transitional regions of the categorical profile space rather than mutually exclusive mathematical clusters. D.3

Quantitative Findings

Silhouette analysis was used to measure the mathematical cohesion of the 9 Functional Families against the raw 6D coordinates. The metrics confirm that functional intent (Goal) does not strictly dictate mechanical structure: • Raw Gower Silhouette (+0.0173): Indicates baseline structural overlap across families in the high-dimensional space. • UMAP Silhouette (−0.0346): Reflects compactness pressure on highly diverse families during lower-dimensional embedding. Furthermore, intra-family distance variance highlights exactly where technical diversity occurs. The Knowledge Modification and Activation-Based family exhibited the highest internal diversity

A Six-Dimensional Taxonomy of Post-Training Adaptation Techniques with Applications in AI Governance

83

(mean intra-family Gower distance = 0.7222). These metrics quantitatively support the visual topology presented in Figure 2, demonstrating how single-axis classification forces artificial boundaries on fundamentally multi-dimensional techniques. D.4

Dimensional Drivers of Topology

Fig. 3. UMAP projection of the latent space, with each panel colored by an individual dimension D1 –D6 .

To isolate how individual axes of our framework influence the global geometric space, we additionally generated a per-dimension topological breakdown (Figure 3). By recoloring the UMAP projection according to each independent dimension, this visualization reveals the precise structural drivers of the latent space. For instance, the clustering of the Inference-Time Adaptation family is mainly governed by a convergence on Mechanism ( d1 = Context Injection ) and Scope ( d5 = Input/Output-Space ). This dimensional breakdown empirically demonstrates that while techniques share primary structural anchors, no single dimension exerts monopolistic control over the adaptation landscape. E

Governance and Compliance Detail

This appendix provides extended treatment of three governance topics that the main text covers in summary form: (i) the lifecycle-stage-by-stage mapping between framework dimensions and the NIST AI RMF; (ii) the three-way breakdown of how D1 –D6 support MDR/IVDR compliance documentation; and (iii) the broader trustworthiness and responsible-adaptation agenda for FM deployments. E.1

Lifecycle-Stage Mapping to the NIST AI RMF

The U.S. NIST AI RMF (AI RMF 1.0) outlines a seven-stage lifecycle for trustworthy AI systems, namely Plan and Design, Collect and Process Data, Build and Use Model, Verify and Validate, Deploy and Use, Operate and Monitor, and Use or Impacted By, across which organizations apply

84

Afdideh, Seoane, and Abtahi

four core functions: Govern, Map, Measure, and Manage [2]. Complementary management-system standards, notably ISO/IEC 42001 (AI management systems) and ISO/IEC 23894 (AI risk management guidance), provide organizational scaffolding for implementing these controls. In 2025 and 2026, this foundational framework was expanded via the “Generative AI Profile” (NIST-AI-600-1) [20] and the April 2026 “Trustworthy AI Profile” for critical infrastructure concept note [187], emphasizing model-specific threat taxonomies that are highly relevant to post-training modifications.

Operate and Monitor. This stage mandates continuous performance tracking and model updating, directly invoking d2 = {Continual Adaptation, Drift Remediation} (Retraining, CL) and the d4 = {Scheduled Permanent, Ad-hoc Permanen or d4 = {Bounded Cumulative, Unbounded Cumulative} persistence categories that these techniques produce. The framework clarifies what “model update” means in compliance terms: not a single operation, but a choice among D1 (Mechanism) with different D4 (Persistence) and D3 (Data Requirements), each carrying distinct risk profiles. Verify and Validate. This stage requires assessments of model validity, reliability, safety, and fairness. d2 = {Alignment, Safety} (RLHF, DPO), d2 = Reasoning (RLVR), and d2 = Robustness (Adversarial Training) all map directly to these requirements. The NIST Trustworthy and Responsible AI guidelines (AI 100-2 E2025) further formalize the requirement to evaluate adversarial threats throughout the model lifecycle [200], which the framework captures through Adversarial Training (Row 19) and its d2 = Robustness classification. Collect and Process Data. This stage addresses data governance, which connects to D3 (Data Requirements). Techniques requiring preference pairs (RLHF, DPO) introduce human annotator governance requirements absent from techniques using only existing labeled data. Techniques requiring verifiable task data (RLVR) introduce different governance considerations: the provenance and correctness of verification oracles must be documented, but human annotator bias is less of a concern. Techniques requiring decentralized data (FL) invoke data sovereignty and crossjurisdictional compliance requirements. The D3 (Data Requirements) dimension makes these data-governance obligations explicit and technique-specific. E.2

EU MDR/IVDR Three-Way Breakdown

Under the EU MDR, software qualifying as a medical device, including AI/ML-driven MDSW, is subject to lifecycle obligations extending beyond initial conformity assessment. Article 83 requires manufacturers to plan, establish, document, implement, maintain, and update a PMS system proportionate to the risk class and appropriate for the device type [59]. Article 84 further requires that this system be based on a PMS plan, which forms part of the technical documentation for non-custom-made devices [59]. The IVDR establishes parallel post-market obligations for in vitro diagnostic devices [60]. Although the MDR does not explicitly distinguish adaptive AI/ML systems, its requirement for continuous post-market data collection, evaluation, and documentation implies ongoing lifecycle documentation obligations, which are particularly demanding for systems that evolve over time. The 6-dimensional framework supports MDR/IVDR compliance documentation in three specific ways. First, D1 (Mechanism) and D4 (Persistence) support technical documentation. The technical-documentation requirement is that the manufacturer describe the design and characteristics of the device, including planned modifications. A statement that a deployed model “was updated” is insufficient; MDR Annex II/III documentation expectations are better met by stating that the model received a d1 = Parametric Update producing a d4 = {Scheduled Permanent, Ad-hoc Permanent} , d5 = {Partial, Modular} change via a specified PEFT method.

A Six-Dimensional Taxonomy of Post-Training Adaptation Techniques with Applications in AI Governance

85

Second, D2 (Goal) and D3 (Data Requirements) support benefit-risk and clinical-evaluation documentation. A change driven by d2 = Drift Remediation using d3 = Small Labeled has a different benefit-risk profile from a change driven by d2 = Task Specialization using the same data category, even if D1 (Mechanism) and D4 (Persistence) are identical. Third, D4 (Persistence) and D6 (Model Type) inform change-control framing. These dimensions support decisions about whether a planned modification is best handled within an approved changecontrol framework, treated as a substantial modification requiring renewed conformity assessment, or addressed through PMS-driven updates to technical documentation. The framework does not make these determinations, which remain context-dependent, but it provides the vocabulary needed to describe the intervention precisely enough that the determination can be made consistently. The MDR’s vigilance reporting obligations (Articles 87–92) and the IVDR’s corresponding obligations (Articles 82–87) are not adaptation events themselves but may trigger them: a reported incident may motivate a d2 = Drift Remediation retraining, a d2 = Remove Knowledge unlearning step, or a d2 = Alignment update. The framework’s D2 (Goal) axis makes the link between vigilance obligations and the appropriate adaptation response explicit. E.3

Trustworthiness and Responsible Adaptation

Beyond specific regulatory frameworks, the broader trustworthiness agenda for FM deployments intersects with adaptation at multiple dimensions. Safety and alignment. The d2 = Alignment category (RLHF, DPO) and d2 = Reasoning category (RLVR) represent the primary technical mechanisms for embedding safety constraints and reasoning capability in LLMs and MLLMs. The distinction among alignment, reasoning, and specialization (all D2 categories) is critical: a model well-specialized for a domain but poorly aligned may produce harmful outputs; a model well-aligned but lacking reasoning depth may fail at complex tasks; a model that reasons well but is unspecialized may produce unhelpful outputs. Governance frameworks must address all three goals, and the D2 (Goal) axis makes this requirement explicit. Data provenance and attribution. The production of synthetic content by FMs introduces requirements for training-data verification and attribution [82]. The D3 (Data Requirements) dimension identifies which techniques consume which data types, enabling traceability from adaptation output back to training-data provenance, a requirement explicitly mandated by emerging clinical AI reporting standards such as TRIPOD+AI for adaptive prediction models [44]. Techniques using preference pairs require documentation of annotator selection, guidelines, and potential biases. Techniques using external corpora (RAG, d3 = External Corpus require documentation of retrieval-source quality and currency. Dynamic benchmarking and continual monitoring. Static evaluation is insufficient for adapted models, a limitation heavily emphasized by the clinical-AI-governance community [70, 112]. Because adaptation may improve performance on a target task while simultaneously degrading safety benchmarks or exacerbating bias, experts emphasize the need for continual, localized monitoring of clinical AI [149]. Dynamic benchmarking platforms that evaluate models across multiple D2 goals simultaneously [96] align with the framework’s multi-dimensional design: a well-adapted model should score well on its target D2 (Goal) without regression on safety-related D2 (Goal). Privacy. The d5 = Distributed category (FL) and PEFT-based personalization address the tension between adaptation and privacy. The framework identifies this tension structurally: personalization ( d2 = Personalization ) requires user data ( d3 = User Interaction Data ), which conflicts with the

86

Afdideh, Seoane, and Abtahi

d2 = Privacy Preservation goal. FL and differential-privacy mechanisms resolve this conflict by operating on d5 = Distributed , ensuring that raw data never leaves the user’s device.

Right to be forgotten and knowledge removal. The GDPR’s right to erasure and similar regulations create a direct mandate for the d2 = Remove Knowledge goal category [1]. Machine Unlearning (Row 47) is the primary technical mechanism for compliance, but its current limitations (approximate rather than exact erasure, vulnerability to relearning attacks) mean that governance frameworks may need to specify acceptable thresholds for knowledge suppression rather than requiring provable deletion. Knowledge editing (Row 48) addresses a related compliance need: correcting factual errors or outdated information without full model replacement. Both techniques intersect with the EU AI Act’s post-market monitoring requirements, since factual errors discovered post-deployment may require targeted correction rather than wholesale retraining.

Related documents

• A configurational pathway study on enhancing the governance level of national fitness public services.#1091280
• Platform Adaptation Under Governance Interventions: Actor Best-Response Modeling and an External Public-Case Benchmark#612558
• Platform Adaptation Under Governance Interventions: Actor Best-Response Modeling and an External Public-Case Benchmark#611081
• Machine learning reveals heterogeneity in the effectiveness of democracy-strengthening interventions#445496
• MANTA: Multi-Agent Network Topology Adaptation for Self-Evolving Multi-Agent Systems#414137