Universal Activation Verbalizer: A Unified Framework for Cross-Model Activation Explanation

Unknown

Conceptio › Archive › arXiv CS

arXiv CSopen access

Universal Activation Verbalizer: A Unified Framework for Cross-Model Activation Explanation

Unknown · 2026 · arxiv_cs

arXiv CS · Papers · License: Open Access · 2026

Open Source ↗Direct PDF ↓

machine learning, deep learning, neural networks

Universal Activation Verbalizer: A Unified Framework for Cross-Model Activation Explanation Haiyan Zhao1

Zirui He1 Guanchu Wang2 Ali Payani3 Yingcong Li1 Mengnan Du4† 1 New Jersey Institute of Technology 2 University of North Carolina at Charlotte 3 Cisco Research 4 The Chinese University of Hong Kong, Shenzhen {hz54,zh296,yingcong.li}@njit.edu [email protected] [email protected] [email protected] † Corresponding author

arXiv:2605.25903v1 [cs.CL] 25 May 2026

Abstract

up view of model internals (Shu et al., 2025; Bricken et al., 2023; Cunningham et al., 2023). Recently, activation verbalization has emerged as a more direct approach: instead of predicting predefined labels or discovering individual features, a language model is trained to describe the information contained in an activation using natural language (Pan et al., 2026; Karvonen et al., 2025). This makes activation explanations more flexible and human-readable. However, existing training-based verbalizers are largely limited to selfexplanation, where a model is trained to explain only its own activations. This restriction makes them difficult to reuse across different model families, architectures, or scales, and can be inefficient when analyzing large donor models. We address this limitation with U NIVERSAL AC TIVATION V ERBALIZER (UAV), which explains activations from heterogeneous donor models with a shared decoder. UAV maps donor activations into decoder-readable soft tokens through a trainable adapter, for which we study both MLPstyle projection (Liu et al., 2023) and Q-Formerstyle attention (Li et al., 2023). The model is trained with activation-to-text alignment followed by explanation-oriented instruction tuning. To improve transferability, we introduce adapter-only transfer, which freezes a decoder-side LoRA (Hu et al., 2022) learned from one donor and trains only a new adapter for another. We evaluate UAV on generative activationgrounded QA across classification, fact retrieval, and gist summarization tasks. UAV remains competitive with strong self-explanation baselines while enabling cross-model verbalization across Llama, Gemma, Yi, and Qwen-family donor models, with larger decoders generally improving performance. Our analyses further show that decoderside tuning improves task-level behavior, while the adapter is essential for recovering input-specific factual and semantic information from activations.

Activation verbalization explains hidden representations in natural language, but existing methods are mostly limited to self-explanation, where each model explains only its own activations. We introduce Universal Activation Verbalizer (UAV), a framework that uses a shared decoder to explain activations from heterogeneous donor models. UAV learns a lightweight adapter that converts donor activations into soft tokens in decoder’s embedding space, and further supports adapteronly transfer by reusing a frozen decoder-side LoRA while training only a new adapter for another donor. Across classification, fact retrieval, and gist summarization, UAV remains competitive with strong self-explanation baselines while enabling cross-model verbalization across model families and scales. Ablations show that decoder-side tuning mainly improves task behavior, whereas the adapter provides the activation-grounded factual and semantic information needed for faithful explanations.

Introduction

Representation learning has been fundamental to the success of large language models (LLMs) (Achiam et al., 2023; Team et al., 2023; Liu et al., 2024). However, understanding what information is encoded in their internal activations remains a central challenge. Prior interpretability work has studied activations from several complementary perspectives. Probing methods test whether linguistic, syntactic, or semantic properties are decodable from hidden states (Belinkov, 2022; Alain and Bengio, 2017), while patchingbased methods inspect representations by transplanting activations into carefully designed decoding contexts (Ghandeharioun et al., 2024; Chen et al., 2024). In parallel, sparse autoencoders and dictionary-learning methods decompose activations into interpretable features, offering a bottom1

⋮

⋮ k

to t1

+ LoRA

Verbalizer Model ℳv

⋮

t 0̂ t 1̂

⋮

Assistant

Update LoRA

t î

Backpropagate to update Adapter

<act> Question

Answer Target Output Tokens)

CE Loss L

rec

Verbalizer Model ℳv

Predicted Answer Tokens

only

y0̂ y1̂

⋮

What does this representation encode?

Reconstructed Prefix Tokens

+ LoRA

Stage 2 Instruction Following Alignment User

⋮

Decoder Blocks ✖ N Frozen Transformer Block

LM Head

⋮

Text / Prompt Tokens

Frozen Transformer Block

Stage 1 Reconstruct Target Tokens

z1 z2

⋮

z1 z2

⋮

⋮ ⋮

⋮

⋮ ⋮

⋮ 0

z1:n ∈ ℝn×de

K/V

z1 z2

Soft Tokens

Q Former Q

⋮ ⋮

⋮ L

+ LoRA

Embedding Layer

MLP

hkℓ ∈ ℝd

⋮

ℓ

Verbalizer Model ℳv

Adapter

⋮

Hidden State at Layer ℓ Position k

⋮

Donor Model ℳ

+ LoRA

yî

CE Loss Lexp

Backpropagate to update Adapter

plain activations directly. Usually, LLMs are employed to verbalize the explanations toward target activations. In training-free paradigm, Patchscopes (Ghandeharioun et al., 2024) and SelfIE (Chen et al., 2024) sought to patch target activations directly into prompts, and utilize LLM to verbalize attributes and semantic information within. Alternatively, LatentQA (Pan et al., 2026) and AO (Karvonen et al., 2025) finetune LLMs to read their own activations from various layers by activation injection. But aforementioned approaches are incapable of cross-model explanations, which make them less generalizable and inefficient.

”

(

Our contributions are threefold: (1) we propose UAV, a unified adapter-decoder framework for cross-model activation verbalization; (2) we introduce adapter-only transfer, enabling reuse of a frozen explanation-oriented decoder while adapting only a donor-specific adapter; and (3) we provide systematic experiments and ablations showing that UAV supports activation explanation across model families and scales while remaining competitive with self-explanation methods. :

𝒜

”

𝒜

Figure 1: Overview of Universal Activation Verbalizer (UAV).

Related Work

Prior works on explaining model activations can be categorized into bottom-up and top-down methods. Bottom-up approaches usually study what information is encoded in activations by defining or discovering interpretable features. For example, probing methods train lightweight classifiers on frozen representations to test whether linguistic, syntactic, or semantic properties are linearly decodable from activations (Belinkov, 2022; Belinkov et al., 2017; Hewitt and Manning, 2019; Rimsky et al., 2024; Conneau and Kiela, 2018; Tenney et al., 2019). Meanwhile, sparse autoencoders (SAEs) perform unsupervised feature discovery by decomposing activations into sparselyactivated monosemantic latent features, which may correspond to concepts at different levels (Huben et al., 2024; Bricken et al., 2023; Cunningham et al., 2023; Achiam et al., 2023). However, this line of research is limited in directly providing humaninterpretable explanations to activations. In contrast, top-down approaches aim to ex-

UAV Framework

We introduce UAV framework that utilizes a trainable adapter to align donor activation space with the decoder model’s embedding space (See Figure 1). In this work, we use representations and activations interchangeably. 3.1

Problem Statement

We study transformer-based LLMs whose last token representations are generally taken as an integration of information from all previous tokens. Give an input text whose input token sequence is ⟨t0 , . . . , tk ⟩, we define activation at position i as hi that encoding all previous i tokens. Our goal is to train an verbalizer Mv to answer questions related to a specific activation hℓi from the ℓ-th layer of donor model M, where two models can be different. These questions can be trivial knowledge within the input sequence or general questions 2

z1 z2

Architecture Design

LayerNorm + Dropout

To improve cross-model alignment, our framework consists of two stages: (1) activation-to-text alignment and (2) instruction-following alignment. In the first stage, we align the donor model’s activation space with the embedding space of the decoder model through input inversion tasks. This stage serves as a warm-up that enables the adapter to translate activations into a representation format that the decoder can effectively consume. In the second stage, we further fine-tune the model so that it can follow explanation-oriented instructions and generate natural-language explanations.

Residual Add Feed Forward Network ( Optional ) Residual Add

q1 q2

⋮

Activation-to-Text Alignment. We freeze the base decoder model and fine-tune only the adapter on input inversion tasks. Given an input token sequence ⟨t0 , . . . , tk ⟩ and the corresponding donor activation hi at position i (0 ≤ i ≤ k), the goal of input inversion is to reconstruct the prefix ⟨t0 , . . . , ti ⟩ from hi under teacher forcing. The formal training objective is provided in Appendix A.1. To perform this activation-to-text mapping efficiently, we compare two adapter architectures: an MLP-based adapter (Liu et al., 2023) and an attention-based adapter following the Q-Former design (Li et al., 2023). The detailed designs of both adapters are provided in Appendix A.2. As shown in Figure 1, the adapter bridges the donor activation space and the decoder embedding space by mapping each donor activation hi into n continuous soft tokens ⟨z 1 , . . . , z n ⟩. These soft tokens are prepended to the textual prompt tokens and injected at the embedding layer, serving as an activationconditioned prefix that guides the frozen decoder to reconstruct the input prefix. We train the adapter in a layer-wise manner, optimizing a separate adapter for activations from each donor layer.

Learnable Query Tokens (one per soft token)

c1 c2

cm Context

⋮

q1 q2

K,V

⋮

Multi-Head Cross Attention

✖L

LayerNorm

⋮

3.2

zn Soft Tokens

⋮

about gist of input text.

Input

Figure 2: Q-Former Architecture

decoder, this stage further refines the activation-totext alignment while improving the model’s ability to follow explanation-oriented instructions. 3.3

Two Training Strategies

To demonstrate the universality of our approach, we consider two training strategies. Full Two-Stage Adaptation. In the first strategy, we use a common decoder backbone to explain activations from different donor models, while performing the full two-stage training procedure separately for each donor as shown in Figure 1. Specifically, for each donor model, we first train a donorspecific adapter for activation-to-text alignment with the frozen decoder. We then jointly fine-tune the adapter and decoder-side LoRA parameters for explanation-oriented instruction following.

Adapter-Decoder Framework. In this stage, we continue training the adapter initialized from the activation-to-text alignment stage, together with the decoder-side trainable parameters, to answer questions about the input sequence. As shown in Figure 1, the second stage replaces the reconstruction objective with instruction-following queries that target information contained in the input text. These queries cover both fine-grained attributes, such as “What was this person’s profession?”, and topic-level understanding, such as “What is the gist of this text?”. By jointly training the adapter and the

Adapter-Only Transfer. In the second strategy, we test whether the instruction-following ability learned from one donor can transfer to others. We first perform full two-stage adaptation on a source donor, then freeze the resulting decoder-side LoRA. For each new donor, we train only a new adapter from scratch while keeping both the decoder backbone and LoRA fixed. This evaluates whether adapter-only training can align unseen donor activations to an already adapted decoder space. 3

Table 1: Adapter architecture exploration. We organize the ablations into two adapter families: MLP-style adapters and Q-Former adapters. For MLP adapters, we vary the bottleneck size and the number of projected soft tokens. For Q-Former adapters, configurations are denoted by (L, C, r/α).

Experiments

In this section, we first study the adapter design used for activation-to-text alignment. We then evaluate UAV along four dimensions: effectiveness against prior methods, generality across donor models and decoder sizes, contributions of each training component, and sensitivity to activation layer, data scale, and activation-to-text warmup. 4.1

Configuration Ada.

LoRA Val ↓ ROUGE-L ↑ BERTS ↑

MLP-style adapters Bottleneck size ID Bottleneck MLP-B1 1024 MLP-B2 2048 MLP-B3 3072 MLP-B4 N/A

Experimental Setup

Models. We use Gemma-3-4B/12B (Google, 2025b,a), Llama-3.1-8B (Meta AI, 2024), Yi-1.534B (01-ai, 2024), and Qwen3-4B (Qwen Team, 2025d) as activation donors. For decoder-side studies, we mainly use Qwen-family decoders, including Qwen3-0.6B/4B/8B/32B (Qwen Team, 2025b,d,e,c) and Qwen2.5-72B (Qwen Team, 2025a). In self-decoding experiments, the donor itself serves as the decoder, providing a naturally aligned reference for cross-model verbalization.

LoRA Val ↓ 132.1M 1.6798 132.1M 1.6818 132.1M 1.6880 132.1M 1.6738

Tok-F1 ↑ 0.258 0.259 0.260 0.296

BERTS ↑ 0.345 0.352 0.35 0.353

Number of projected soft tokens ID n Ada. LoRA Val ↓ MLP-T1 8 99.7M 132.1M 1.6833 MLP-T2 16 162.6M 132.1M 1.6921 MLP-T3 64 540M 132.1M 1.6880 MLP-T4 128 1044M 132.1M 1.7407

Tok-F1 ↑ 0.265 0.259 0.260 0.250

BERTS ↑ 0.354 0.352 0.350 0.341

Tok-F1 ↑ 0.267 0.272 0.278 0.284

BERTS ↑ 0.351 0.352 0.357 0.369

Ada. 195M 367M 540M 860M

Q-Former adapters ID (L, C, r/α) Ada. LoRA Val ↓ QF-Base (1, 8, 16/32) 162.8M 33.0M 1.6818 QF-Deep (4, 8, 16/32) 398.8M 33.0M 1.6863 QF-Wide (4, 16, 16/32) 482.7M 33.0M 1.6755 QF-LoRA (2, 8, 64/128) 241.4M 132.1M 1.6388

Note. For Q-Former adapters, L denotes the number of QFormer layers, C denotes the context width, and r/α denotes the decoder-side LoRA rank and scaling factor. “Ada.” denotes the number of trainable adapter parameters.

Datasets. We use generative question answering as the downstream task. Our training corpus is a 17-source mixture covering factual, scientific, topical, affective, classification-style, and conversational inputs, including Wikipedia entity snippets, peS2o scientific abstracts (Soldaini and Lo, 2023), AG News (Zhang et al., 2015), SST-2 (Socher et al., 2013), datasets reused from Activation Oracle (Karvonen et al., 2025), LMSYS-Chat-1M user turns (Zheng et al., 2024), and dair-ai/emotion tweets (Saravia et al., 2018). All examples are tokenized with the Qwen3-4B tokenizer and capped at 64 tokens to match the activation extraction window. The resulting corpus contains 469K Stage 1 samples and 950K Stage 2 QA pairs, with full composition and preprocessing details in Appendix B.

overlap metrics and semantic similarity metrics. Specifically, ROUGE-L, chrF++, and token-level F1 measure sequence-level, character-level, and lexical overlap with the reference, respectively, while BERTScore captures semantic similarity using contextualized embeddings. Higher scores indicate stronger agreement between the generated answer and the reference text. Detailed definitions of all metrics are provided in Appendix C. 4.2

Adapter Architectures for Activation-to-Text Alignment

To understand how adapter architecture affects UAV performance, we conduct separate design studies for MLP-style and Q-Former-style adapters. The result is shown in Table 1. Full results are detailed in Appendix D.1

Baselines. We compare with both training-free and training-based activation verbalization methods. PatchScopes (Ghandeharioun et al., 2024) and SelfIE (Chen et al., 2024) are representative training-free approaches, while LatentQA (Pan et al., 2026) and Activation Oracle (AO) (Karvonen et al., 2025) are latest training-based methods. We evaluate these baselines in the same generative QA setting, using reconstruction- and QA-oriented metrics to assess how well each method recovers information encoded in activations.

MLP-style adapters. For MLP-style adapters, we vary the bottleneck size and the number of projected soft tokens while keeping the LoRA setup fixed. Increasing the bottleneck dimension from 1024 to 3072 does not improve performance, and removing the bottleneck achieves the best validation loss among MLP variants. This suggests that bottleneck compression may discard information useful for activation inversion.

Metrics. Since the questions are open-ended, we evaluate generated answers from complementary perspectives, using both token- or character-level 4

Table 2: Main comparison across classification, fact retrieval, and gist summarization tasks. We report ROUGE-L and BERTScore in the main paper as high-level generation quality metrics. Method

Decoder

Injected Layer

Classification R-L

BERTS

Fact R-L

BERTS

Gist R-L

BERTS

Overall R-L

BERTS

Donor: Qwen3-4B-Instruct-2507 UAV Qwen3-4B (Full) UAV Qwen3-4B (AOT from Llama) AO Qwen3-4B LatentQA Qwen3-4B SelfIE Qwen3-4B PatchScope Qwen3-4B

0 0 1 1 3 1

0.272±0.146 0.315±0.164 0.233±0.327 0.374±0.345 0.273±0.177 0.323±0.200 0.254±0.252 0.344±0.271 0.276±0.143 0.323±0.157 0.239±0.335 0.372±0.350 0.281±0.191 0.326±0.212 0.260±0.258 0.347±0.274 0.270±0.147 0.319±0.154 0.129±0.195 0.158±0.287 0.239±0.152 0.270±0.186 0.198±0.181 0.234±0.242 0.264±0.135 0.308±0.171 0.200±0.313 0.349±0.338 0.269±0.175 0.309±0.193 0.235±0.243 0.327±0.266 0.106±0.047 -0.064±0.092 0.019±0.034 -0.343±0.135 0.101±0.052 -0.087±0.117 0.064±0.060 -0.199±0.180 0.109±0.049 -0.047±0.090 0.018±0.035 -0.340±0.133 0.103±0.053 -0.077±0.115 0.065±0.062 -0.190±0.183

Donor: Llama-3.1-8B-Instruct UAV Llama-8B (Full) UAV Qwen3-4B (Full) UAV Qwen3-4B (AOT from Qwen) AO Llama-8B LatentQA Llama-8B SelfIE Llama-8B PatchScope Llama-8B

0 0 0 1 1 3 1

0.286±0.138 0.356±0.155 0.295±0.369 0.421±0.368 0.264±0.166 0.317±0.187 0.286±0.275 0.379±0.283 0.289±0.146 0.353±0.159 0.265±0.348 0.402±0.356 0.270±0.179 0.319±0.207 0.274±0.265 0.369±0.278 0.277±0.135 0.321±0.160 0.240±0.335 0.381±0.351 0.266±0.180 0.308±0.212 0.257±0.255 0.346±0.276 0.288±0.142 0.344±0.155 0.263±0.352 0.397±0.353 0.271±0.182 0.312±0.203 0.273±0.267 0.362±0.275 0.291±0.146 0.348±0.160 0.280±0.367 0.408±0.365 0.284±0.178 0.317±0.193 0.285±0.276 0.370±0.282 0.101±0.063 -0.006±0.100 0.022±0.050 -0.285±0.166 0.113±0.102 -0.005±0.169 0.066±0.080 -0.136±0.204 0.120±0.056 0.016±0.083 0.017±0.035 -0.322±0.147 0.114±0.086 -0.009±0.135 0.071±0.075 -0.147±0.207

Note. All entries are reported as mean with standard deviation shown in subscript. F ULL denotes full two-stage adaptation on the current donor–decoder pair: Stage 1 trains the adapter with the decoder frozen, and Stage 2 jointly tunes the adapter and decoder-side LoRA while keeping the decoder backbone frozen. AOT denotes adapter-only transfer. For Qwen3-4B (AOT from Llama), we reuse and freeze the decoder-side LoRA obtained from the F ULL adaptation of Llama-8B donor to Qwen3-4B decoder, and train only a new adapter for the Qwen3-4B donor. For Llama-8B (AOT from Qwen), we reuse and freeze the decoder-side LoRA obtained from the F ULL adaptation of Qwen3-4B donor to Llama-8B decoder, and train only a new adapter for the Llama-8B donor.

We also find that increasing the number of projected soft tokens does not yield monotonic gains. The 128-token variant produces twice as many soft tokens as the maximum input length, but has the largest adapter size and the worst validation loss. Together, these results indicate that adding bottlenecks or excessive soft tokens does not naively improve verbalization, and too many soft tokens may introduce redundancy or noise and make optimization harder. Therefore, we favor setting the softtoken length close to the maximum input length.

We adopt QF-LoRA as the default setting for subsequent experiments, as it achieves the lowest validation loss and the strongest semantic similarity while maintaining better adapter-side parameter efficiency than the strongest MLP variant. 4.3 Comparison with Activation Verbalization Baselines We evaluate UAV against representative activation verbalization baselines under a shared generative QA protocol covering classification, fact retrieval, and gist summarization. These tasks require recovering information at different levels from a single donor activation, from label-level decisions to finegrained facts and high-level summaries; examples are provided in Appendix B.3. We compare against training-free patching methods, including SelfIE and PatchScope, as well as training-based selfexplanation methods, including AO and LatentQA. For training-based baselines, we use their native self-explanation setting, which is favorable because no cross-model alignment is required. Thus, this comparison tests whether UAV can remain competitive with specialized self-explanation methods while supporting cross-model verbalization. As shown in Table 2, under the self-explanation setting, UAV achieves strong performance against prior training-based approaches. For Qwen3-4B, our full two-stage adaptation obtains the best

Q-Former adapters. For Q-Former adapters, we vary the number of cross-attention layers, context slots, and decoder-side LoRA capacity. Increasing depth from QF-Base to QF-Deep only yields marginal gains in generation metrics and does not reduce validation loss, showing that deeper adapterside transformation alone is insufficient. Increasing context width from QF-Deep to QF-Wide gives more consistent improvements, indicating that a wider latent memory helps preserve activation information. The strongest result is achieved by QFLoRA, which increases decoder-side LoRA capacity while using fewer adapter parameters than QFDeep and QF-Wide. This suggests that, after the adapter aligns donor activations with the decoder embedding space, decoder-side adaptation is more important for converting the inverted activation signal into natural-language outputs. 5

Table 3: Cross-donor verbalization with Qwen3-4BInstruct-2507 as the shared decoder. We report validation loss, ROUGE-L, and BERTScore in the main paper. Donor

Val ↓

R-L ↑

Table 4: Effect of decoder size on cross-decoding performance. All results use Llama-3.1-8B-Instruct layer-27 activations as the donor and Qwen-family decoders of different sizes. Decoder

BERTS ↑

Val ↓ Tok-F1 ↑ R-L ↑ chrF ↑ BERTS ↑

Llama-3.1-8B-Instruct 1.6388 0.274±0.265 0.369±0.278 Gemma-3-4B-IT 1.6790 0.266±0.263 0.354±0.274 Gemma-3-12B-IT 1.6589 0.269±0.258 0.355±0.281 Yi-1.5-34B-Chat 1.5837 0.299±0.277 0.377±0.296

Qwen3-0.6B 1.9172 Qwen3-4B 1.6388 Qwen3-8B 1.6263 Qwen3-14B 1.5050

Note. All rows use Qwen3-4B-Instruct-2507 as the shared decoder. Val denotes validation loss, R-L denotes ROUGEL, and BERTS denotes BERTScore. The best result per column is highlighted in bold.

Note. All rows use Llama-3.1-8B-Instruct layer-27 activations as the donor. Val denotes the best fine-tuning validation loss; R-L denotes ROUGE-L; BERTS denotes BERTScore. The best result in each column is highlighted in bold.

overall ROUGE-L and BERTScore among selfexplaining methods, outperforming both AO and LatentQA. The gain is especially clear on fact retrieval, where UAV achieves substantially higher BERTScore than AO, suggesting that the learned adapter-decoder alignment better preserves finegrained factual information encoded in the activation. For Llama-3.1-8B, our full Llama-8B verbalizer achieves the strongest overall BERTScore and remains competitive with LatentQA on ROUGEL, while outperforming AO on fact retrieval and overall semantic similarity. These results show that UAV is not merely a cross-model extension, but also a strong activation verbalizer in the conventional self-explanation setting. Qualitative examples in Table 11 further illustrate successful factual recovery while highlighting metric brittleness and typical failure cases for short answers. UAV shows a larger advantage over training-free patching methods. SelfIE and PatchScope obtain substantially lower scores across both donors, especially on fact retrieval, where their BERTScore values are near zero or negative. This indicates that directly patching activations into prompts can recover limited topical cues but struggles with finegrained factual and semantic information. By learning to map donor activations into decoder-readable soft tokens, UAV enables the decoder to use the activation signal more reliably. Finally, UAV demonstrates strong transferability through adapter-only transfer. In this setting, the decoder-side LoRA learned from one donor is frozen, and only a new adapter is trained for the current donor. Despite this restricted training setup, AOT remains close to full two-stage adaptation. For Qwen3-4B, AOT even slightly outperforms the corresponding full self-explanation setting on the overall metrics. For Llama-3.1-8B, AOT shows

0.250 0.284 0.231 0.309

0.239 0.274 0.222 0.296

0.248 0.281 0.228 0.300

0.323 0.369 0.315 0.388

only a moderate drop compared with full adaptation while still clearly outperforming training-free baselines and remaining competitive with trainingbased self-explanation methods. These results highlight a key advantage of UAV: it can reuse decoderside explanation behavior across donors and adapt to a new donor by training only the adapter. 4.4

Model-Agnostic Verbalization Across Target Models

We study UAV’s model-agnosticity by varying donor models and decoder sizes, and compare cross-model verbalization with self-decoding. Shared Decoder Across Donor Models. We further test whether a shared decoder can verbalize activations from different donor models. As shown in Table 3, we fix Qwen3-4B-Instruct-2507 as the decoder and vary the donor across Llama, Gemma, and Yi families. UAV achieves stable verbalization performance across all tested heterogeneous donors, showing that the adapter-decoder interface is not tied to a single model family. The results also suggest that decoding difficulty is not determined by model size alone. While Gemma-3-12B improves over Gemma-3-4B, Llama-3.1-8B outperforms both Gemma donors, and Yi-1.5-34B achieves the best overall performance. These comparisons indicate that crossdonor verbalization depends not only on donor scale, but also on donor-specific factors that affect alignment to the shared decoder space. Effect of Decoder Size. We study how decoder capacity affects cross-decoding by fixing the donor as Llama-3.1-8B-Instruct layer-27 activations and varying the Qwen-family decoder size. As shown in Table 4, increasing the decoder from 0.6B to 4B substantially improves both validation loss and 6

ROUGE-L

Token F1 Comp

Fact

Gist

Comp 0.329

0.318

0.328

0.3

Gist 0.297

0.3

BERTScore 0.309

0.308

0.4

0.265

Comp

Fact

Gist 0.321

0.262

0.179

0.171 0.147

0.141 0.112

0.1

0.114

0.143

0.162 0.144

0.098

0.104

0.015

0.0

Rand. No LoRA Rand FT-LoRA + base + LoRA only

−0.4

Base

Full UAV

0.045

0.0

Rand. No LoRA Rand FT-LoRA + base + LoRA only

Ablation variant

Full UAV

0.345

0.402 0.361 0.282

0.196 0.168

0.017

-0.019 -0.045

−0.2

0.056 0.018

0.015

0.0 Base

0.145 0.141

0.120 0.095

0.058 0.019

0.155 0.141

0.145 0.123

0.1

0.104

0.190 0.151

0.2

0.207

0.2

Score

0.2

Score

0.233

Score

Fact

-0.195 -0.235

-0.164

-0.333 -0.403

Base

Rand. No LoRA Rand FT-LoRA + base + LoRA only

Full UAV

Ablation variant

Figure 3: Disentangling the roles of the activation adapter and decoder-side LoRA. Token F1 0.30

0.297

ROUGE-L 0.295

0.284

0.286

0.28

0.284

0.28

0.29

0.276

0.280

0.274

0.27

Llama-3.1-8B Gemma-3-4B Gemma-3-12B

0.270

0.269

0.266

Llama-3.1-8B Gemma-3-4B Gemma-3-12B

chrF

BERTScore

0.30

0.293

0.290

0.28

chrF, and BERTScore, showing that the shared Qwen3-4B decoder can still recover substantial information from non-Qwen activations. The gap is smallest for Gemma-3-4B, where cross-model decoding even slightly improves BERTScore, suggesting that the external decoder can sometimes provide comparable or better semantic matching despite lower lexical overlap. Task-level comparisons are provided in Appendix D.4

0.285

0.38

0.379

0.377 0.369

0.281

0.280 0.276

0.272

Llama-3.1-8B Gemma-3-4B Gemma-3-12B

Self-decoding

0.36

0.354 0.350

0.355

Llama-3.1-8B Gemma-3-4B Gemma-3-12B

Qwen3-4B (cross)

4.5

Figure 4: Overall performance comparison between self-decoding and cross-model verbalization.

Validating the Role of Activation Signal and Decoder Models

We further disentangle the roles of the activation adapter and the decoder-side fine-tuning in UAV. To this end, we compare several controlled variants: (1) BASE, which uses the base decoder without an adapter or decoder-side LoRA; (2) R AND .+BASE, which injects randomly initialized adapter outputs into the base decoder; (3) S TAGE -1, which uses the adapter trained only by activation-to-text alignment with the base decoder; (4) R AND . A LIGN, which trains with randomly mismatched activationtarget pairs; (5) FT-L O RA ONLY, which keeps the Stage-2 fine-tuned decoder-side LoRA but removes the activation-conditioned adapter outputs; and (6) F ULL UAV, which uses both the trained adapter and the fine-tuned decoder-side LoRA with correctly aligned activations. Figure 3 shows that the base decoder alone performs poorly, especially on fact retrieval. This indicates that the evaluation questions cannot be answered reliably from the decoder prior alone. Adding a randomly initialized adapter to the base decoder does not improve the results, confirming that arbitrary soft-token injection cannot substitute for meaningful activation alignment. The Stage-1 adapter provides a modest improvement, showing that activation-to-text alignment helps the decoder access some information from donor activations, although Stage-1 alignment alone remains insufficient for downstream explanation-oriented QA.

generation quality. The 14B decoder achieves the best performance across all metrics, suggesting that larger decoders can better convert aligned activation signals into faithful natural-language answers. However, the trend is not strictly monotonic: Qwen3-8B obtains a slightly lower validation loss than Qwen3-4B but performs worse on all generation metrics. This indicates that decoder scaling alone does not guarantee better activation verbalization, and that optimization dynamics or decoderspecific instruction-following behavior can also affect the final generation quality. External Decoder vs. Self-Decoding We compare cross-model verbalization with self-decoding, where each donor also serves as its own decoder. As shown in Figure 4, self-decoding generally achieves stronger results across most donors and metrics, suggesting that a model’s own decoder is naturally better aligned with its internal activation space. In contrast, using Qwen3-4B as a shared external decoder requires the adapter to bridge heterogeneous representation spaces, making crossmodel verbalization more challenging. Despite this disadvantage, the performance gap remains moderate. For Llama-3.1-8B and Gemma3-12B, cross-model verbalization retains competitive overall scores across Token-F1, ROUGE-L, 7

0.38

Token-F1 RougeL chrF++ BERTScore best

0.325 0.300

Score ↑

0.275

Score

Token-F1 RougeL chrF++ BERTScore best

0.300

0.250

15 21 Layer

(a) Layerwise

0.340

0.32 0.30

0.297

0.302 0.291

0.28

200

422 750 QA slots (k)

0.26

951

(b) Data Scale

Tok-F1

0.288 0.279

ROUGE-L chrF BERTScore Generation metric

(a) 500K

0.374

w/o Stage-1 + Stage-1

0.36

0.354

0.34

0.284

0.275

0.225

0.38

w/o Stage-1 + Stage-1

0.36

0.350

Score ↑

0.375 0.325

0.356

0.34 0.321

0.32 0.30

0.308

0.306

0.298 0.287

0.285

0.28 0.26

Tok-F1

ROUGE-L chrF BERTScore Generation metric

(b) 951K

Figure 5: Additional ablation results on activation layer Figure 6: Effect of Stage-1 pretraining on generation and Stage 2 training data scale. (a) Layer-wise verbal- metrics. We compare w/o Stage-1 and +Stage-1 under izer performance using activations from different layers (a) 500k and (b) 951k Stage-2 training data. Stage-1 of Qwen3-4B. (b) Effect of Stage 2 QA data scale on pretraining consistently improves all generation metrics, verbalization performance. especially under the full-data setting.

Training Data Scale. We study the effect of Stage 2 data scale by training with 200K, 422K, 750K, and 951K query-oriented QA pairs. Figure 5b shows that larger and more diverse QA supervision improves downstream activation interpretation, so we use 951K QA pairs as the default scale. Additional training curves and per-metric results are provided in Figure 11.

The comparison between FT-L O RA ONLY and F ULL UAV further reveals complementary roles of decoder-side adaptation and the activation adapter. The fine-tuned LoRA alone substantially improves comprehension-style performance, suggesting that Stage-2 decoder tuning helps the model learn the task format and classification-oriented decision patterns from the QA supervision. However, FTL O RA ONLY remains much weaker on fact retrieval and gist summarization, where the model must recover input-specific information from the donor activation. In contrast, F ULL UAV achieves the strongest performance on these informationintensive tasks, demonstrating that the trained adapter provides essential activation-conditioned evidence beyond what is captured by decoder-side LoRA or question priors. These results suggest that decoder-side fine-tuning mainly improves instruction following and task-level behavior, while the adapter is critical for extracting fine-grained factual and semantic content from activations. 4.6

Activation-to-Text Warmup. We further study the role of the Stage-1 activation-to-text alignment stage. Stage-1 serves as a warmup that first aligns donor activations with the decoder embedding space before query-oriented fine-tuning. As shown in Figure 6, removing this warmup consistently hurts performance across both training scales, while also increasing validation loss and perplexity. The improvement is more pronounced under the full training setting, suggesting that Stage1 provides a better initialization for subsequent activation-grounded QA. These results show that activation-to-text alignment is an important prerequisite for reliable verbalizer training.

Additional Ablations

Activation Layer. We evaluate how the choice of activation layer affects verbalization performance by decoding activations from different depths of the donor model. Specifically, under the selfexplanation setting, we study Layers 3, 9, 15, 21, 27, and 33 of Qwen3-4B. As shown in Figure 5a, performance generally improves from early to middle layers, with the strongest overall results around the middle-to-late layers. This suggests that these layers provide more informative and interpretable representations for activation verbalization, consistent with prior observations that later layers tend to encode richer and more abstract semantic knowledge (Jin et al., 2025). Additional layer-wise ablation results are provided in Appendix D.7.

Conclusions and Future Work

We present U NIVERSAL ACTIVATION V ERBAL IZER (UAV), an adapter-decoder framework that explains activations from heterogeneous donor models with a shared decoder. UAV remains competitive with self-explanation baselines while enabling cross-model verbalization across model families and scales. Ablations show that decoder-side tuning improves task behavior, whereas the adapter provides activation-grounded factual and semantic information. Future work will improve fact retrieval and gist understanding through stronger activation alignment, multi-layer features, and richer factual or summarization supervision. 8

Limitations

Hoagy Cunningham, Aidan Ewart, Logan Riggs, Robert Huben, and Lee Sharkey. 2023. Sparse Autoencoders Find Highly Interpretable Features in Language Models. arXiv preprint. ArXiv:2309.08600 [cs].

Although UAV enables cross-model activation verbalization, several limitations remain. First, UAV remains limited in fact retrieval and gist summarization, especially when recovering fine-grained attributes or abstract semantic content from a single activation. Second, our current setting focuses on short-context inputs, leaving long-context and document-level activation explanation for future work. Third, adapter-only transfer still requires training a donor- and layer-specific adapter, which may be costly for large-scale layer-wise analysis.

Asma Ghandeharioun, Avi Caciularu, Adam Pearce, Lucas Dixon, and Mor Geva. 2024. Patchscopes: a unifying framework for inspecting hidden representations of language models. In Proceedings of the 41st International Conference on Machine Learning, pages 15466–15490. Google. 2025a. Gemma-3-12b. https: //huggingface.co/google/gemma-3-12b-it. Google. 2025b. Gemma-3-4b. https://huggingface. co/google/gemma-3-4b-it. John Hewitt and Christopher D Manning. 2019. A structural probe for finding syntax in word representations. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4129–4138.

References 01-ai. 2024. Yi-1.5-34b-chat. https://huggingface. co/01-ai/Yi-1.5-34B-Chat. Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, and 1 others. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774.

Edward J Hu, yelong shen, Phillip Wallis, Zeyuan AllenZhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations.

Guillaume Alain and Yoshua Bengio. 2017. Understanding intermediate layers using linear classifier probes.

Robert Huben, Hoagy Cunningham, Logan Smith, Aidan Ewart, and Lee Sharkey. 2024. Sparse autoencoders find highly interpretable features in language models. In International Conference on Learning Representations, volume 2024, pages 7827–7845.

Yonatan Belinkov. 2022. Probing classifiers: Promises, shortcomings, and advances. Computational Linguistics, 48(1):207–219. Yonatan Belinkov, Nadir Durrani, Fahim Dalvi, Hassan Sajjad, and James Glass. 2017. What do neural machine translation models learn about morphology? In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 861–872.

Mingyu Jin, Qinkai Yu, Jingyuan Huang, Qingcheng Zeng, Zhenting Wang, Wenyue Hua, Haiyan Zhao, Kai Mei, Yanda Meng, Kaize Ding, and 1 others. 2025. Exploring concept depth: How large language models acquire knowledge and concept at different layers? The 31st International Conference on Computational Linguistics (COLING 2025).

Trenton Bricken, Adly Templeton, Joshua Batson, Brian Chen, Adam Jermyn, Tom Conerly, Nick Turner, Cem Anil, Carson Denison, Amanda Askell, Robert Lasenby, Yifan Wu, Shauna Kravec, Nicholas Schiefer, Tim Maxwell, Nicholas Joseph, Zac Hatfield-Dodds, Alex Tamkin, Karina Nguyen, and 6 others. 2023. Towards monosemanticity: Decomposing language models with dictionary learning. Transformer Circuits Thread. Https://transformercircuits.pub/2023/monosemanticfeatures/index.html.

Adam Karvonen, James Chua, Clément Dumas, Kit Fraser-Taliente, Subhash Kantamneni, Julian Minder, Euan Ong, Arnab Sen Sharma, Daniel Wen, Owain Evans, and 1 others. 2025. Activation oracles: Training and evaluating llms as general-purpose activation explainers. arXiv preprint arXiv:2512.15674. 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, pages 19730–19742.

Haozhe Chen, Carl Vondrick, and Chengzhi Mao. 2024. Selfie: self-interpretation of large language model embeddings. In Proceedings of the 41st International Conference on Machine Learning, pages 7373–7388.

Chin-Yew Lin. 2004. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74–81.

Alexis Conneau and Douwe Kiela. 2018. Senteval: An evaluation toolkit for universal sentence representations. In Proceedings of the eleventh international conference on language resources and evaluation (LREC 2018).

Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, and 1 others.

2024. Deepseek-v3 technical report. arXiv preprint arXiv:2412.19437.

Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D. Manning, Andrew Ng, and Christopher Potts. 2013. Recursive deep models for semantic compositionality over a sentiment treebank. In Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 1631–1642. Association for Computational Linguistics.

Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. 2023. Visual instruction tuning. Advances in neural information processing systems, 36:34892– 34916. Meta AI. 2024. Llama-3.1-8b. https://huggingface. co/meta-llama/Llama-3.1-8B. Alexander Pan, Lijie Chen, and Jacob Steinhardt. 2026. LatentQA: Teaching LLMs to decode activations into natural language. In The Fourteenth International Conference on Learning Representations.

Gemini Team, Rohan Anil, Sebastian Borgeaud, JeanBaptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, and 1 others. 2023. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805.

Maja Popović. 2017. chrF++: words helping character n-grams. In Proceedings of the Second Conference on Machine Translation, pages 612–618. Association for Computational Linguistics.

Ian Tenney, Patrick Xia, Berlin Chen, Alex Wang, Adam Poliak, R Thomas McCoy, Najoung Kim, Benjamin Van Durme, Sam Bowman, Dipanjan Das, and Ellie Pavlick. 2019. What do you learn from context? probing for sentence structure in contextualized word representations. In International Conference on Learning Representations.

Qwen Team. 2025a. Qwen2.5-72b. https:// huggingface.co/Qwen/Qwen2.5-72B-Instruct. Qwen Team. 2025b. Qwen3-0.6b. huggingface.co/Qwen/Qwen3-0.6B.

https://

Qwen Team. 2025c. Qwen3-32b. huggingface.co/Qwen/Qwen3-32B.

https://

Qwen Team. 2025d. Qwen3-4b. huggingface.co/Qwen/Qwen3-4B.

https://

Qwen Team. 2025e. Qwen3-8b. huggingface.co/Qwen/Qwen3-8B.

https://

Luca Soldaini and Kyle Lo. 2023. peS2o (Pretraining Efficiently on S2ORC) Dataset. Technical report, Allen Institute for AI. ODC-By, https://github. com/allenai/pes2o.

Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q Weinberger, and Yoav Artzi. 2020. Bertscore: Evaluating text generation with bert. The International Conference on Learning Representations (ICLR). Xiang Zhang, Junbo Zhao, and Yann LeCun. 2015. Character-level convolutional networks for text classification. Advances in neural information processing systems, 28.

Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. 2016. SQuAD: 100,000+ questions for machine comprehension of text. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 2383–2392, Austin, Texas. Association for Computational Linguistics.

Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Tianle Li, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zhuohan Li, Zi Lin, Eric Xing, and 1 others. 2024. Lmsys-chat-1m: A large-scale real-world llm conversation dataset. In International Conference on Learning Representations, volume 2024, pages 22225– 22257.

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), pages 15504–15522. Elvis Saravia, Hsien-Chi Toby Liu, Yen-Hao Huang, Junlin Wu, and Yi-Shin Chen. 2018. CARER: Contextualized affect representations for emotion recognition. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 3687–3697. Association for Computational Linguistics. Dong Shu, Xuansheng Wu, Haiyan Zhao, Daking Rai, Ziyu Yao, Ninghao Liu, and Mengnan Du. 2025. A survey on sparse autoencoders: Interpreting the internal mechanisms of large language models. EMNLP Findings.

Additional Method Details

A.1

Teaching Forcing

MLP-base design. Given an input activation h ∈ RdDonor , the MLP adapter first applies layer normalization and projects the activation into a hidden space. By default, we set the hidden dimension of z 1 to d1 = 2dDonor , unless an explicit bottleneck dimension b is used.

We use teacher forcing for both the activation-totext alignment stage and the explanation-oriented instruction-tuning stage. Given an input token sequence t0:k = ⟨t0 , . . . , tk ⟩ and a donor activation hℓi from layer ℓ at position i, the adapter first maps the activation into n decoder-readable soft tokens z1:n = Aϕ (hℓi ). During Stage 1, the target output is the prefix t0:i . At decoding step j, instead of conditioning on previously generated tokens t̂<j , we feed the ground-truth prefix t<j to the verbalizer. The conditional likelihood is therefore pθ,ϕ (t0:i | hℓi ) =

i Y

pθ,ϕ (tj | z1:n , t<j ) ,

h̃ = LN(h), z 1 = Dropout GELU(W1 h̃ + b1 ) , z b = Dropout(GELU(Wb z 1 + bb )) , y = Wout z ⋆ + bout , Y = reshape(y; n × d). Here, the bottleneck transformation that produces z b is optional. When the bottleneck layer is disabled, we set z ⋆ = z 1 and d⋆ = d1 ; otherwise, we set z ⋆ = z b and d⋆ = b. The projection matrices are W1 ∈ Rdh ×dDonor , Wb ∈ Rb×dh , and Wout ∈ R(n·d)×d⋆ .

(1)

j=0

where ϕ denotes adapter parameters and θ denotes the trainable decoder-side parameters. In Stage 1, θ is frozen and only ϕ is optimized. The Stage 1 teacher-forcing objective is the token-level negative log-likelihood:

Attention-layer-based design. Given an input activation h ∈ RdDonor , the cross-attention adapter follows a Q-Former-style design. It first expands the donor activation into a set of latent context slots, and then uses learnable query tokens to extract information from these slots through cross-attention. The final query states are used as output soft tokens in the decoder embedding space. We first project the donor activation into M context slots in the decoder embedding space. These slots serve as a latent memory that stores different views of the donor activation.

Lrec = −

1 X log pθ,ϕ (tj | z1:n , t<j ) . (2) i+1 j=0

For Stage 2, each training example contains a textual prompt or question x and a target answer sequence y1:T = ⟨y1 , . . . , yT ⟩. The soft tokens are prepended to the prompt embeddings, and the verbalizer predicts the answer autoregressively. Under teacher forcing, the model conditions on the ground-truth answer prefix y<j at each decoding step: pθ,ϕ (y1:T | hℓi , x) =

T Y

C = LNc (reshape (Wctx h + bctx ; M × d)) ,

pθ,ϕ (yj | z1:n , x, y<j ) .

(6)

where C ∈ RM ×d denotes the constructed context slots, M is the number of context slots, and d is the decoder embedding dimension. The projection matrix is Wctx ∈ R(M d)×dDonor , with bias bctx ∈ RM d . Intuitively, this step maps a single donor activation vector into a sequence of M latent tokens, which can later be attended to by the learnable queries. After constructing the context slots, we introduce n learnable query tokens. Each query token corresponds to one output soft token.

j=1

(3) The explanation-oriented loss is T

1X Lexp = − log pθ,ϕ (yj | z1:n , x, y<j ) . (4) T j=1

Thus, teacher forcing replaces the modelgenerated history with the corresponding groundtruth history during training. This stabilizes optimization by exposing the adapter and decoder to correct prefixes at every step, while inference remains fully autoregressive. A.2

(5)

Q(0) = Q0 ,

Adapter Designs

Q0 ∈ Rn×d .

(7)

Here, n is the number of output soft tokens. Unlike the context slots C, which are input-dependent, Q0 is a learnable parameter shared across inputs.

We adopted two different adapter designs to study which is more efficient under the setting of UAV. 11

Through cross-attention, these queries become conditioned on the current donor activation. We stack L cross-attention blocks. At layer ℓ, the query states Q(ℓ−1) attend to the context slots C. The cross-attention sublayer uses prenormalization and a residual connection. Q̂(ℓ) = LNq Q(ℓ−1) ,

adapter, this design decouples the number of output tokens n from the main projection parameters: increasing n mainly enlarges the learnable query matrix Q0 , while the context projection and attention parameters remain unchanged.

Datasets

Ĉ = LNc (C),

The detailed combinations of datasets are shown in Table 5.

Q′(ℓ) = Wq Q̂(ℓ) ,

[K (ℓ) , V (ℓ) ] = Wkv Ĉ, (ℓ) (ℓ) (ℓ) Qh , Kh , Vh = split-heads Q′(ℓ) , K (ℓ) , V (ℓ) , !! (ℓ) (ℓ) ⊤ Qh Kh (ℓ) √ A = Dropout softmax , dh (ℓ) O(ℓ) = merge-heads A(ℓ) Vh , (ℓ) (ℓ−1) Q̃ = Q + Dropout Wo O(ℓ) . (8)

B.1

Data Curation. To train a universal activation verbalizer, we construct a diversified short-context corpus that covers factual knowledge, semantic comprehension, classification-oriented attributes, and natural user intents. Our goal is not to train a long-document reader, but to expose the verbalizer to diverse semantic phenomena under a controlled input length budget where most sequences are within 64 tokens. The full data construction pipeline consists of raw text sourcing, length normalization, train/evaluation split construction, offline QA generation, and activation representation caching.

We use H attention heads and set dh = d/H. (ℓ) After splitting into heads, Qh ∈ RH×n×dh and (ℓ) (ℓ) Kh , Vh ∈ RH×M ×dh . Therefore, the attention map satisfies A(ℓ) ∈ RH×n×M , which indicates how each query token attends to the M context slots. The projection matrices are Wq ∈ Rd×d , Wkv ∈ R2d×d , and Wo ∈ Rd×d . Each cross-attention block can optionally include a feed-forward sublayer. When r = ffn_mult > 0, we update the query states as

Raw Text Sources. We collect raw texts from 17 sources spanning three regimes. First, we use Wikipedia entries to cover factual knowledge, grouping pages into seven subtypes: person, place, event, organization, work, concept, and generic. Second, we include scientific abstracts, news, sentiment, emotion, gender, namedentity, and LatentQA-style control datasets to cover short-context semantic and classification-oriented attributes. Third, we sample natural user requests from LMSYS-Chat-1M by retaining only the first user turn from each conversation after language, moderation, length, PII, and deduplication filtering. All examples are normalized into a unified schema containing the text, source, subtype, and metadata fields.

Q(ℓ) = Q̃(ℓ) + Dropout W2 GELU W1 LN Q̃(ℓ) , (9)

where W1 ∈ Rrd×d and W2 ∈ Rd×rd . We set ffn_mult = 4 for all our Q-Former-based adapters, as we found that the FFN sublayer substantially improves the performance of our verbalizer. When ffn_mult = 0, the FFN sublayer is disabled and we set Q(ℓ) = Q̃(ℓ) .

(10)

After L cross-attention blocks, the final query states are normalized and used as the adapter output: Y = Dropout LNout Q(L) , Y ∈ Rn×d .

Data Sourcing

Length Normalization. We normalize all raw texts into short activation contexts before caching representations or generating QA supervision. For longer examples, we count tokens with the Qwen34B tokenizer, which is also the target model used for activation caching. If an example exceeds 64 tokens, we first truncate it to the leading sentence. The sentence splitter searches for real sentence boundaries while skipping common abbreviations such as “Dr.”, “U.S.”, and “Ph.D.”. If no reliable

(11)

In this design, the context slots C are inputdependent and provide a latent memory of the donor activation, whereas the query tokens Q0 are input-independent learnable slots that define the output token positions. Through cross-attention, each query token learns to retrieve information from the context slots, resulting in activationconditioned soft tokens. Compared with the MLP 12

Table 5: Composition of the diversified training corpus. All 17 sources are used for Stage 1 activation-to-text reconstruction, while the 15 sources with synthesized QA pairs are also used for Stage 2 generative QA finetuning. Texts denotes the number of source passages retained after sentence packing, length filtering (10–64 Qwen3-4B tokens), and deduplication. QA items denotes the number of flattened (activation, Q, A) training triples derived from LLM-generated QA records; “-” indicates sources used only in Stage 1. Source

Domain

Texts QA mode

QA items

Stage 1 (alignment / activation–text reconstruction): 17 sources Wikipedia (first-paragraph snippets, factual QA) Person Biography Place Geography Event Events Organization Institutions Work Books / films Concept Abstract concepts Generic Miscellaneous

7,250 7,250 7,250 7,250 7,250 7,250 7,250

factual factual factual factual factual factual factual

27,898 27,773 27,808 27,887 27,892 27,888 27,830

Long-form and classification (comprehension QA) peS2o (Scientific) Sci. abstracts AG News News headlines SST-2† Movie reviews TweetEval-Sentiment† Tweets TweetEval-Emotion† Tweets dair-ai/emotion Tweets LatentQA-Control† Diverse short text MD-Gender† Gendered text NER† Entity spans

50,250 50,250 18,254 20,250 3,074 19,912 16,490 29,857 14,923

comprehension comprehension comprehension comprehension comprehension comprehension comprehension -

186,408 199,707 62,876 78,225 11,015 77,521 64,662 -

User queries (intent / query QA) LMSYS-Chat-1M (1st user turn) Chatbot requests

20,000 query

Stage-1-only deduplicated supplements (no synthetic QA) AG News (extra) News headlines 50,000 TweetEval-Sentiment (extra) Tweets 25,299 dair-ai/emotion (extra) Tweets 50,000 LMSYS-Chat-1M (extra) Chatbot requests 50,000 Stage 1 alignment total (17 sources)

75,326 -

469,309

Stage 2 (generative QA finetune): 15 sources ⊆ Stage 1 Stage 2 QA finetune total 244,662

950,716

†

Sources labeled as reused are drawn from the Activation Oracle corpus (Karvonen et al., 2025). For each source, the final min(250, 20%) rows are held out for evaluation, and reported counts include this evaluation tail. The four extra blocks are disjoint samples from the corresponding upstream corpora, with all Stage-2 texts removed via hash- or index-based deduplication to prevent leakage between Stage 1 activation–text alignment and Stage 2 QA finetuning. QA items are generated offline using Qwen3-14B served by vLLM.

sentence boundary is found, we fall back to the leading paragraph prefix. The truncated text is then re-validated, and examples that still violate the 30character or 64-token constraint are discarded.

remains bounded by 64 Qwen3-4B tokens. For LMSYS user queries, we additionally apply a sourcespecific character filter before global normalization, keeping only first-turn user requests between 30 and 400 characters and removing redaction artifacts that start with NAME_.

Some sources require a wider parsing window to extract metadata, but this window is never used as model input. For example, Wikipedia examples are parsed with a longer lead window to identify entity descriptors and assign factual subtypes, while classification-style sources may use a wider window to infer labels or entity spans. These metadata fields are stored only as prompt-side hints for QA generation. The actual text field still comes from the globally normalized record and

This design makes the activation cache predictable and keeps QA generation lightweight. Since all cached inputs are token-bounded under the same tokenizer as the target model, representation shards have a stable memory footprint. Moreover, because every input text is at most 64 tokens, the full QA-generation prompt remains well below the vLLM context limit, even after adding source13

Table 6: Source distribution of the test set. Each source contributes 100 examples. Non-Wikipedia sources contain comprehension-style and gist questions, while Wikipedia sources contain fact-retrieval questions. Category

Source

Topic / news

ag_news

–

100

Sentiment

sst2 tweeteval_sentiment

28 25

72 75

– –

100 100

Emotion

tweeteval_emotion dair_emotion

25 26

75 74

– –

100 100

26 25 27

74 75 73

– – –

100 100 100

Wikipedia

wikipedia_concept wikipedia_event wikipedia_generic wikipedia_organization wikipedia_person wikipedia_place wikipedia_work

– – – – – – –

100 100 100 100 100 100 100

Total

15 sources

207

593

700

1500

User intent lmsys_user Behavior / control latentqa_control Scientific abstract scientific

straints: questions must refer to the subject only by a generic placeholder ("this person", "this place", "this entity", etc.) rather than by its proper name; questions must not exceed 15 words; and questions must not quote four or more consecutive words from the passage. For the remaining subtypes we use a comprehension mode that yields one one-sentence gist together with one to three QA pairs covering higherlevel aspects such as the main event, topic, intent, tone, sentiment, emotion, or speaker attitude. Each comprehension subtype carries a custom extra_hints field that names the dimensions of interest — for instance, aspects praised or criticized, overall verdict for SST-2, trigger or object of the feeling, overall mood for the emotion sources, and subject of debate, what the speaker argues for TweetEvalstance. The prompt explicitly bans generic template questions ("What is the overall tone?", "What is the intent of the text?") and requires answers to be expressed through concrete, content-grounded phrases lifted from the passage. For labeled subtypes (SST-2, AG News, TweetEval-sentiment/emotion/stance, DAIRemotion, MD-Gender, NER, LatentQA-control) the categorical label is supplied to the generator only as a focus hint identifying which dimension at least one QA pair must address — e.g., the speaker’s attitude or evaluative tone for SST-2, the subject area or field of the news reported for AG News, the emotional state expressed by the speaker for DAIR-emotion, the position taken on the subject for stance, and the specific named entity mentioned in the text for NER. The prompt forbids the answer from copying the label string or close synonyms, requiring instead that the relevant property be described through distinctive phrases from the passage. This design reduces label leakage and prevents the adapter from learning a shortcut that maps activations directly to dataset labels. Comprehension-mode examples are augmented with gist supervision. The gist answer is generated once as a standalone one-sentence paraphrase; at training time it is paired with a question sampled uniformly from a pool of 51 semantically equivalent summarization prompts (e.g., "What is the main idea of this text?", "Summarize this text in one sentence.", "State the gist of this text."). This question diversification prevents the adapter from overfitting to a fixed query template and encourages stable behavior under surface-level prompt

Gist Comp. Fact Total

specific instructions or label-focus hints. Overall, QA generation, activation caching, and dataloader joining operate on the same normalized short-text distribution. B.2

QA Pair Construction.

Given each passage, we generate question–answer pairs that supervise the adapter on different facets of the source content. We construct a corpus of approximately 838K passages spanning 17 subtypes drawn from heterogeneous sources: seven Wikipedia subtypes (person, place, event, concept, organization, work, generic), scientific abstracts, AG News, SST-2, TweetEval (sentiment, emotion, stance), DAIR-emotion, MD-Gender, an NER corpus, and a LatentQA control set. All passages are truncated to ≤64 tokens (Qwen tokenizer) and filtered to ≥30 characters so that every input fits a uniform context budget. Each passage is routed to one of two generation modes according to its subtype. For Wikipedia passages we use a factual mode in which the generator produces up to four atomic QA pairs, each targeting a single attribute. A subtypespecific fact_hint directs the generator toward attribute classes that are typical for the entity type — for example, nationality, profession, dates, affiliations, relationships for person; founding date, location, purpose, leadership for organization; creator, year, genre, cast, setting, reception for work; and dates, participants, locations, causes, outcomes for event. To prevent trivially solvable questions, the prompt additionally enforces three anti-leak con14

Table 7: Adapter architecture exploration. We organize the ablations into two adapter families: MLP-style adapters and Q-Former adapters. For MLP adapters, we vary the bottleneck size and the number of projected soft tokens. For Q-Former adapters, configurations are denoted by (L, C, r/α).

variation. All QA pairs are generated by Qwen3-14BAWQ served through vLLM (batch size 32, max new tokens 300, JSON-constrained decoding). The full corpus is partitioned into uniform 20K-row shards across subtypes ( 42 shards) and each shard’s raw model outputs are persisted to disk before parsing, so that prompt or schema revisions can be applied offline without re-running generation. B.3

Concrete Examples for QA Tasks

Q-Former adapters QF-Base (1, 8, 16/32) QF-Deep (4, 8, 16/32) QF-Wide (4, 16, 16/32) QF-LoRA (2, 8, 64/128)

P =

(12)

0.276 0.271 0.272 0.260

0.271 0.266 0.268 0.259

0.267 0.272 0.278 0.321

0.261 0.268 0.273 0.306

Pchar + Pword Rchar + Rword , R= . (15) 2 2

chrF++ is then computed as:

ROUGE-L. ROUGE-L measures sequence-level similarity based on the longest common subsequence (LCS) between the generated answer and the reference (Lin, 2004). Given a generated sequence x and a reference sequence y, let LCS(x, y) denote the length of their longest common subsequence. The LCS-based recall and precision are:

chrF++ =

(1 + β 2 )P R , β2P + R

(16)

where β controls the relative weight of recall and precision. Following the standard setting, we use character n-grams up to order 6, word n-grams up to order 2, and β = 2. BERTScore. BERTScore measures semantic similarity by comparing contextualized token embeddings of the generated answer and the reference (Zhang et al., 2020). Given generated tokens n x = {xi }m i=1 and reference tokens y = {yj }j=1 , BERTScore first computes pairwise cosine similarities between their embeddings. Precision and

LCS(x, y) RLCS = , |y|

LCS(x, y) PLCS = . |x| (13) ROUGE-L is then computed as: (1 + β 2 )RLCS PLCS , RLCS + β 2 PLCS

0.264 0.267 0.268 0.297

chrF++. chrF++ computes an F -score over character n-grams and word n-grams (Popović, 2017). Let Pchar and Rchar denote the average precision and recall over character n-grams, and let Pword and Rword denote the corresponding averages over word n-grams. The combined precision and recall are:

Token-level F1. Token-level F1 measures lexical overlap between the generated answer and the reference at the token level, following the standard QA evaluation protocol (Rajpurkar et al., 2016). Given precision P and recall R over shared tokens, the score is computed as:

ROUGE-L =

162.8M 33.0M 398.8M 33.0M 482.7M 33.0M 241.4M 132.1M

0.270 0.269 0.272 0.308

where β controls the relative importance of recall and precision.

We use five automatic metrics to evaluate the agreement between generated answers and references.

2P R . P +R

LoRA Tok-F1 ↑ chrF++ ↑

Note. For Q-Former adapters, L denotes the number of Q-Former layers, C denotes the context width, and r/α denotes the decoder-side LoRA rank and scaling factor. “Ada.” denotes the number of trainable adapter parameters.

Evaluation Metrics

F1 =

Ada.

MLP-style adapters Bottleneck size MLP-B1 1024 195M 132.1M MLP-B2 2048 367M 132.1M MLP-B3 3072 540M 132.1M MLP-B4 N/A 860M 132.1M Number of projected soft tokens MLP-T1 8 99.7M 132.1M MLP-T2 16 162.6M 132.1M MLP-T3 64 540M 132.1M MLP-T4 128 1044M 132.1M

Table 10 summarizes the composition of our test set. We sample 100 examples from each of the 15 sources, resulting in 1,500 evaluation examples in total. For non-Wikipedia sources, we construct both comprehension-style questions and gist summarization questions. For Wikipedia-based sources, we focus on fact-retrieval questions that ask for finegrained attributes from the input text. This design allows the evaluation set to cover three complementary levels of activation-grounded understanding: task-oriented comprehension, high-level gist summarization, and fine-grained factual retrieval.

Configuration

(14) 15

Table 8: Additional comparison across classification, fact retrieval, and gist summarization tasks. We report Token-F1 and chrF as complementary token-level and character-level generation metrics. Method

Injected Layer

Decoder

Classification Tok-F1

chrF

Fact Tok-F1

Gist chrF

Tok-F1

Overall chrF

Tok-F1

chrF

Donor: Qwen3-4B-Instruct-2507 UAV Qwen3-4B (Full) UAV Qwen3-4B (AOT from Llama) AO Qwen3-4B LatentQA Qwen3-4B SelfIE Qwen3-4B PatchScope Qwen3-4B

0 0 1 1 3 1

0.291±0.148 0.271±0.122 0.229±0.326 0.244±0.289 0.297±0.181 0.283±0.152 0.263±0.254 0.261±0.221 0.298±0.145 0.276±0.118 0.236±0.334 0.252±0.293 0.301±0.195 0.285±0.165 0.269±0.260 0.267±0.224 0.292±0.135 0.279±0.109 0.131±0.197 0.187±0.188 0.259±0.157 0.256±0.123 0.210±0.186 0.231±0.159 0.284±0.140 0.266±0.108 0.199±0.312 0.225±0.278 0.290±0.176 0.275±0.144 0.246±0.245 0.249±0.211 0.122±0.056 0.187±0.053 0.020±0.038 0.053±0.047 0.117±0.065 0.176±0.060 0.073±0.071 0.122±0.083 0.126±0.059 0.190±0.054 0.019±0.038 0.055±0.047 0.120±0.066 0.176±0.062 0.075±0.074 0.124±0.084

Donor: Llama-3.1-8B-Instruct UAV Llama-8B (Full) UAV Qwen3-4B (Full) UAV Qwen3-4B (AOT from Qwen) AO Llama-8B LatentQA Llama-8B SelfIE Llama-8B PatchScope Llama-8B

0 0 0 1 1 3 1

0.307±0.147 0.283±0.120 0.294±0.368 0.303±0.334 0.288±0.175 0.271±0.144 0.297±0.277 0.290±0.247 0.309±0.150 0.284±0.124 0.262±0.349 0.281±0.313 0.293±0.182 0.277±0.155 0.284±0.267 0.281±0.236 0.298±0.139 0.273±0.113 0.239±0.335 0.257±0.299 0.285±0.180 0.275±0.158 0.268±0.257 0.266±0.226 0.311±0.145 0.285±0.120 0.262±0.351 0.277±0.318 0.293±0.185 0.275±0.154 0.285±0.268 0.279±0.238 0.313±0.150 0.285±0.124 0.280±0.367 0.296±0.334 0.306±0.182 0.280±0.151 0.296±0.278 0.289±0.249 0.123±0.081 0.164±0.062 0.022±0.053 0.074±0.063 0.128±0.109 0.176±0.088 0.077±0.092 0.124±0.083 0.156±0.080 0.192±0.058 0.018±0.039 0.066±0.050 0.138±0.099 0.182±0.078 0.088±0.096 0.131±0.085

recall are then defined by greedy token matching:

projected soft tokens: using more soft tokens does not monotonically improve performance, and the largest setting with 128 tokens even leads to lower scores, despite using more than one billion trainable adapter parameters. This indicates that overparameterized MLP projections may introduce redundancy or optimization difficulty rather than consistently improving activation-to-text alignment. For Q-Former adapters, increasing depth and context width provides moderate gains over the shallow baseline, but the most effective configuration is obtained by increasing the decoder-side LoRA capacity. Specifically, QF-LoRA achieves the best Tok-F1 and chrF++ scores while using substantially fewer adapter parameters than the largest MLP variants. This suggests that the attentionbased adapter is more parameter-efficient for transforming donor activations into decoder-readable representations, and that sufficient decoder-side adaptation is important for exploiting these representations. Based on these results, we adopt the Q-Former adapter with enhanced LoRA capacity as our default architecture.

PBERT =

1 X max cos(xi , yj ), j m

1 RBERT = n

i=1 n X

(17)

max cos(xi , yj ).

j=1

The final score is their F1: FBERT =

2PBERT RBERT . PBERT + RBERT

(18)

Because it compares contextualized embeddings rather than exact tokens, BERTScore can reward semantically similar paraphrases with different surface forms.

Additional Experimental Results

D.1

Additional Adapter Architecture Results

Table 7 reports additional adapter architecture ablations with token-level and character-level generation metrics. For MLP-style adapters, increasing the bottleneck size generally improves chrF++, with the no-bottleneck variant achieving the highest chrF++ score among MLP adapters. However, the trend is less consistent for Tok-F1, suggesting that simply increasing the MLP capacity does not always lead to uniformly better lexical alignment. A similar pattern appears when varying the number of

D.2

Additional Comparison with Activation Interpretation Baselines

Table 8 provides complementary token-level and character-level results for the baseline comparison in Section 4.3. While the main paper re16

Table 9: Additional cross-donor verbalization results with Qwen3-4B-Instruct-2507 as the shared decoder. We report Token-F1 and chrF as complementary lexical and character-level metrics.

ports ROUGE-L and BERTScore as high-level generation quality metrics, here we additionally report Token-F1 and chrF to evaluate lexical and character-level agreement with the reference answers. The results are consistent with the main comparison in Table 2. For Qwen3-4B activations, UAV achieves the strongest overall performance under both full two-stage adaptation and adapteronly transfer. In particular, AOT from Llama slightly improves over the corresponding full selfexplanation setting on both Token-F1 and chrF, showing that the frozen decoder-side LoRA learned from another donor can still provide an effective explanation-oriented decoder space. Compared with training-based self-explanation baselines, UAV obtains clear gains on fact retrieval, where recovering precise information from the activation is especially important. For Llama-3.1-8B activations, the full Llama-8B verbalizer achieves the best overall Token-F1 and chrF among UAV variants and remains competitive with LatentQA. When using Qwen3-4B as an external decoder, UAV still maintains strong performance, demonstrating that the proposed adapterdecoder framework can verbalize activations across different model families rather than relying only on self-explanation. The adapter-only transfer setting incurs only a moderate drop from full adaptation while still substantially outperforming training-free baselines. Across both donors, SelfIE and PatchScope perform much worse than trained verbalizers, especially on fact retrieval. This further supports the observation that directly patching activations into prompts is insufficient for open-ended activationgrounded QA. In contrast, UAV learns an explicit adapter that maps donor activations into decoderreadable soft tokens, leading to stronger lexical, character-level, and semantic agreement with the reference answers. D.3

Donor

Tok-F1 ↑

chrF ↑

Llama-3.1-8B-Instruct 0.284±0.267 0.281±0.236 Gemma-3-4B-IT 0.276±0.264 0.272±0.231 Gemma-3-12B-IT 0.280±0.261 0.280±0.230 Yi-1.5-34B-Chat 0.310±0.278 0.306±0.250 Note. All rows use Qwen3-4B-Instruct-2507 as the shared decoder. The best result in each column is highlighted in bold.

same family can expose representations that are easier for the verbalizer to decode. However, the comparison across families again shows that scale alone does not fully determine decoding quality: Llama-3.1-8B remains stronger than both Gemma models despite being smaller than Gemma-3-12B. Together with the main results, these additional metrics show that the cross-donor behavior is stable across semantic, lexical, and character-level evaluation. UAV can adapt the same Qwen3-4B decoder to activations from multiple donor families, while the remaining performance variation reflects both donor scale and family-specific representation structure. D.4

Task-level comparison between self-decoding and external-decoder verbalization

Figure 7 provides task-level results for classification, fact retrieval, and gist summarization. The overall trend is consistent with the main-paper comparison: self-decoding usually performs better than cross-model verbalization, indicating that a familymatched decoder benefits from a more compatible activation space. However, the size of the gap varies across task types and donor models. The difference is most consistent on fact retrieval. Across the fact-oriented subset, self-decoding almost always outperforms the Qwen3-4B crossmodel decoder, especially for Llama-3.1-8B and Gemma-3-12B. This suggests that fine-grained factual recovery is more sensitive to representationspace mismatch than classification or gist summarization. Since fact retrieval requires the verbalizer to recover specific entity-level or attribute-level information from the activation, even small alignment errors between the donor activation space and the shared decoder embedding space can lead to degraded lexical and semantic agreement.

Additional Results on Cross Model Decoding

Table 9 reports complementary Token-F1 and chrF results for the cross-donor setting. The trends are consistent with the main-paper metrics in Table 9. Yi-1.5-34B achieves the strongest lexical and character-level agreement, obtaining the best Token-F1 and chrF among all donor models. Gemma-3-12B also improves over Gemma-3-4B, further suggesting that larger models within the 17

Token F1

ROUGE-L

0.31

0.311

0.309 0.307

0.28

0.30 0.28

0.262

0.26

0.265

0.261 0.251

0.251

0.26

0.255

0.297

0.293

Llama-3.1-8B Gemma-3-4B Gemma-3-12B

chrF

BERTScore

chrF

BERTScore

chrF

0.36 0.285

0.284 0.283

0.280

0.279

0.28 0.274

0.356 0.353

0.350

0.34

0.333

0.32

Llama-3.1-8B Gemma-3-4B Gemma-3-12B

Self-decoding

0.30

0.42 0.281

0.28

0.327

0.304

0.303

0.274

0.271

0.423

0.421

0.402 0.387 0.387

0.265

0.315

Llama-3.1-8B Gemma-3-4B Gemma-3-12B

Qwen3-4B (cross)

Self-decoding

(a) Classification

0.28 0.389

Llama-3.1-8B Gemma-3-4B Gemma-3-12B

0.277

0.26

Llama-3.1-8B Gemma-3-4B Gemma-3-12B

BERTScore

0.30

0.33

0.292

0.288

0.40

0.271

0.264

Llama-3.1-8B Gemma-3-4B Gemma-3-12B

0.29

0.274

0.270

0.293

0.288

0.256

0.284

0.283

0.28

0.30 0.261

0.311

0.310

0.286

0.28

ROUGE-L

0.32

0.295

0.284

0.277

0.276 0.277

Token F1

ROUGE-L

0.30 0.294 0.291

0.289 0.286 0.300

0.2990.300

0.30

0.27

Token F1

0.30

0.32

0.277

0.32

0.278

0.271

0.324 0.317

0.319

0.321

0.311

0.31

Llama-3.1-8B Gemma-3-4B Gemma-3-12B

Self-decoding

Qwen3-4B (cross)

(b) Fact

0.311

Llama-3.1-8B Gemma-3-4B Gemma-3-12B

Qwen3-4B (cross)

Figure 7: Task-level comparison between self-decoding and cross-model verbalization on classification, fact retrieval, and gist summarization. 0.30

0.32

Classification Gist Fact

0.30

0.24

Classification Gist Fact

0.24 0.22 0.6B

0.26 Classification Gist Fact

0.22

4B 8B Decoder size

14B

0.6B

BERTScore

0.28 0.26

chrF++

0.28 RougeL

Token-F1

0.30

0.28 0.26

14B

0.35

0.30

0.24

4B 8B Decoder size

Classification Gist Fact

0.40

0.6B

4B 8B Decoder size

14B

0.6B

4B 8B Decoder size

14B

Figure 8: Task-level effect of decoder size on cross-decoding performance. All results use Llama-3.1-8B-Instruct layer-27 activations as the donor and vary the Qwen-family decoder size.

Together, these task-level results clarify the trade-off observed in the main paper. Self-decoding provides the strongest performance when modelspecific decoder adaptation is available, particularly for fact retrieval. Cross-model verbalization, however, offers a more universal setup by using a shared decoder across different donor families, while preserving competitive performance on coarse-grained and semantic explanation tasks.

chrF Comp

Fact

Gist

0.3

0.294

0.307

0.231

Score

0.212

0.2

0.301 0.281

0.226 0.199 0.175 0.165

0.166 0.148 0.148

0.175 0.164

0.108

0.1 0.057 0.041

0.0 Base

Rand. No LoRA Rand FT-LoRA + base + LoRA only

Full UAV

Ablation variant

D.5

Figure 9: Additional chrF results for disentangling the roles of the activation adapter and decoder-side LoRA.

Task-level Performance across Decoder sizes

Figure 8 provides task-level results for the decodersize study in Section 4.4. All settings use Llama3.1-8B-Instruct layer-27 activations as the donor and vary the Qwen-family decoder size. The tasklevel results are broadly consistent with the overall results in the main paper: increasing the decoder from 0.6B to 4B improves most metrics, while Qwen3-8B does not consistently outperform Qwen3-4B despite having comparable validation loss. Qwen3-14B achieves the strongest performance in most task-level settings, especially on fact retrieval and gist summarization. The effect of decoder scaling is not uniform across tasks. Classification is relatively stable across decoder sizes, suggesting that coarse labellevel information is easier to recover once the activation has been mapped into the decoder space. In contrast, fact retrieval is more sensitive to decoder capacity, as it requires recovering fine-grained

For classification and gist summarization, the comparison is more mixed. Classification relies more on coarse label-level signals, while gist summarization mainly requires topic-level semantic information. These types of information appear more robust under cross-model transfer, and in several cases the Qwen3-4B cross-model decoder matches or slightly exceeds self-decoding on individual metrics. For example, cross-model verbalization improves BERTScore for Gemma-3-4B in classification and improves several gist metrics for Gemma-family donors. These results suggest that cross-model verbalization is more effective when the target query depends on high-level semantic information, but remains more challenging for precise factual extraction. 18

1.8

1.00

1.0

0.75

0.5

0.50

best

0.25

2000 Step

4000

best

0.25

0.00 0

0.50

15 21 Layer

0.50 best 0.25

0.00 3

BERTScore

1.00 0.75 RougeL

2.0

1.00 0.75 chrF++

Loss

L27 L33 val train

Token-F1

L3 L9 L15 L21

2.2

best

0.0 −0.5

0.00 3

15 21 Layer

Figure 10: Additional layer-wise results under the self-explanation setting. We report training dynamics and evaluation results for Qwen3-4B activations from Layers 3, 9, 15, 21, 27, and 33. The verbalizer can learn from all evaluated layers, while middle-to-late layers generally yield stronger decoding performance.

1.6

1000 2000 3000 4000 Step

1.00

1.0

0.75

0.5

0.50

best

0.50

best

0.50

0.25

0.00

200

500 750 QA slots (k)

fullqa

200

500 750 QA slots (k)

fullqa

best

BERTScore

1.8

1.00 0.75

chrF++

Token-F1

Loss

fullqa val train

RougeL

200k 500k 750k

2.0

best

0.0 −0.5 −1.0

200

500 750 QA slots (k)

fullqa

200

500 750 QA slots (k)

fullqa

Figure 11: Additional training data scale results. We compare verbalizers trained with 200K, 422K, 750K, and 951K Stage-2 QA pairs. Training remains stable across all scales, and evaluation performance generally improves as the amount of query-oriented supervision increases. The full 951K setting achieves the strongest overall performance across token-level and semantic metrics.

input-specific information from the donor activation. Gist summarization also benefits from the larger decoder, indicating that stronger decoders can better convert aligned activation signals into high-level semantic descriptions. These results suggest that decoder capacity helps activation verbalization, but scaling alone does not guarantee monotonic improvements; decoder-specific optimization and instruction-following behavior also affect final generation quality. D.6

information-intensive tasks, further confirming that the trained adapter is necessary for extracting activation-grounded factual and semantic evidence. D.7

Additional Layer-wise Results.

We provide additional layer-wise results to complement the analysis in Section 4.6. Specifically, we evaluate Qwen3-4B activations from Layers 3, 9, 15, 21, 27, and 33 under the self-explanation setting. Figure 10 reports both the training dynamics and token-level evaluation results. The training and validation losses decrease steadily across all examined layers, suggesting that the verbalizer can learn meaningful activation-to-text mappings from different depths. Meanwhile, the token-level metrics show a mild but consistent advantage for middleto-late layers, with Layer 27 achieving the best Token-F1 and chrF++ scores among the evaluated layers. These additional results are consistent with the observation in the main text that later activations tend to be more informative for verbalization, while early-layer activations are comparatively less effective.

More Results on Disentangling Adapter and Decoder Contributions

Figure 9 provides additional chrF results for the ablation study in Section 4.5, which disentangles the roles of the activation adapter and decoder-side LoRA. The trend is consistent with the Token-F1, ROUGE-L, and BERTScore results in Figure 3. The base decoder and randomly initialized adapter variants perform poorly, showing that the model cannot answer activation-grounded questions from the decoder prior or arbitrary soft-token injection alone. Decoder-side LoRA improves classificationstyle performance, suggesting that Stage-2 tuning helps the decoder learn the QA format and task-level decision patterns. However, LoRA-only variants remain much weaker on fact retrieval and gist summarization, where the answer depends more heavily on input-specific information encoded in the donor activation. In contrast, Full UAV obtains the strongest chrF scores on these

D.8

Additional Training Data Scale Results

We provide additional results on the effect of Stage2 training data scale in Figure 11. Specifically, we train the verbalizer with 200K, 422K, 750K, and 951K QA pairs and report both training dynamics and token-/semantic-level evaluation results in Figure 11. The training curves show stable optimization across different data scales. Meanwhile, 19

the evaluation results indicate that increasing the number of QA pairs consistently improves verbalization quality, especially for the full 951K setting. These results further support the observation in Section 4.6 that larger and more diverse query-oriented supervision helps the verbalizer better interpret activations under downstream queries.

Table 10: Representative examples from the test set. For each source, we show the input text from which activations are extracted, together with the corresponding evaluation question and reference answer. Non-Wikipedia sources use comprehension-style questions, while Wikipedia sources use fact-retrieval questions. Source

Input

ag_news

Mortaza enjoys moment to remember. [Comp.] What event is described in the Mortaza dismissing Sachin Tendulkar Bangladesh fast bowler Mashrafe bin Mortaza text? with the first ball of the day in Chitwas understandably elated after dismissing tagong. Sachin Tendulkar with the first ball of the day in Chittagong. could easily be called the best korean film of [Comp.] What overall evaluation is ex- It could easily be called the best Korean 2002 pressed about the film? film of 2002. “and Then the hilarious Sheamus Cash-In sce- [Comp.] What tone does the text con- The text uses “hilarious” to describe the nario, which must have been a botch on Randy vey about the event described? scenario, indicating a lighthearted and Orton’s part... RAW” humorous tone. @user I don’t think your a girls girl fraud bully [Comp.] What kind of tone does the The tone is confrontational and dismiscelebeffer text convey? sive, as seen in terms like “fraud” and “bully.” i was feeling rather cranky cos i was thinking [Comp.] What is the trigger for the The lack of sleep. about the lack of sleep i had bah speaker’s current emotional state? What are the origins of the name “NAME_1”? [Comp.] What is the main task re- The user is asking for the origins of the quested by the user? name “NAME_1”. In 1985, the Internal Revenue Service (IRS) [Comp.] What is the purpose of the To assist recipients and auditors in imissued IRS Circular A-128, “Audits of State document mentioned in the text? plementing the new Single Audit. and Local Governments,” to assist recipients and auditors in implementing the new Single Audit. With the rapid development of information [Comp.] What field is the text dis- The rapid development of information science and technology, the demand for com- cussing in relation to the growing de- science and technology. puter data processing is increasing, resulting mand for SSDs? in the rapid growth of the demand for highpower and high-performance solid-state drives (SSDs).

sst2 tweeteval_sentiment

tweeteval_emotion

dair_emotion lmsys_user latentqa_control

scientific

wikipedia_concept

wikipedia_event

wikipedia_generic

wikipedia_organization

wikipedia_person

wikipedia_place wikipedia_work

Question

Answer

Erigeron pulchellus, the Robin’s plantain, blue [Fact] What family is this concept part Asteraceae. spring daisy or hairy fleabane, is a North Amer- of? ican species of plants in the family Asteraceae. Adin David Ross (born October 11, 2000) is an [Fact] What type of video games does NBA 2K and Grand Theft Auto V. American online streamer. He is known for his this person stream? collaborations with celebrities and livestreams of the NBA 2K and Grand Theft Auto V video games. The Walter Field House is a historic resi- [Fact] What is the name of this historic The Walter Field House. dence located along Reading Road in northern residence? Cincinnati, Ohio, United States. Ranger Up is an American apparel company [Fact] Who owns and operates this or- Military veterans. that is owned and operated by military veter- ganization? ans. Their T-shirt designs typically involve themes from the armed forces, mixed martial arts, and historical figures. The 1960 South Carolina United States Senate [Fact] Who was the incumbent senator Strom Thurmond. election was held on November 8, 1960 to in this election? select the U.S. Senator from the state of South Carolina. Popular incumbent Senator Strom Thurmond easily won the Democratic primary and was unopposed in the general election. Bacqueville is a commune in the Eure depart- [Fact] What country is this place lo- France. ment in Normandy in northern France. cated in? Handle with Care is a 1985 Australian televi- [Fact] Who directed this work? Paul Cox. sion film directed by Paul Cox and starring Luncinda Cowden and Anna Maria Monticelli. The screenplay concerns two women with breast cancer.

Table 11: Qualitative examples grouped by task category and ROUGE-L score level. For each category and level (Best / Middle / Irrelevant), we show two representative samples drawn from different source datasets.

Classification

Category Level

R-L

Dataset

Best

0.800 ag_news

Best

0.667 tweeteval_emotion

Middle

0.308 ag_news

Middle

0.276 dair_emotion

Irrelevant 0.000 tweeteval_sentiment

Reference

Prediction

What significant deci- EU agrees to membership The European Union The European Union desion did the European talks with Turkey. . . agreed to open member- cided to open formal talks Union make regarding ship talks with Turkey. with Turkey. Turkey? What is the primary emo- I wanna kill you and de- The text conveys a strong The text conveys a strong tion conveyed in the stroy you. I want you died sense of hostility and de- sense of anger and frustratext? and I want Flint back. emo sire for destruction, as seen tion, as seen in phrases like scene fuck die hatered in phrases like “I wanna “kill you” and “I’m so ankill you and destroy you” gry.” and “I want you died.” Which entities are the Syria Redeploys Some Syria and Lebanon are the The main actors are the main actors involved in Security Forces in main actors, with Syria Iraqi army and the U.S.-led the event? Lebanon. . . being the one taking ac- coalition. tion and Lebanon being the country where the security forces are being redeployed. What is the main feel- i wasnt feeling sociable i The speaker is not feel- The speaker feels “so ing expressed by the really wasnt ing sociable and explicitly much more relaxed” and speaker? states, “i wasnt feeling so- “so much more at ease.” ciable i really wasnt.” The text discusses the speaker’s experience of watching the movie “Paper Towns” for the second time. The speaker is beginning to feel more comfortable with the idea of being a “big brother” to the baby. animated television series

Best

What type of media is Camp Lakebottom is a animated television series this work? Canadian animated television series. . . 1.000 wikipedia_organization What was the name of Westfield Corporation was Westfield Group the group this entity split founded with the spin-off from? of the Westfield Group in 2014. . .

Middle

0.400 wikipedia_place

Middle

0.333 wikipedia_generic

Best Fact

Input

What is the main subject Kim Kardashian has been Kim Kardashian looking of the text? looking back fondly on back fondly on 2015. 2015. On Sunday the siren named her top magazine covers of the year. What is the speaker be- im feeling more comfort- Derby ginning to feel more able with derby i feel as comfortable with? though i can start to step out my shell

Irrelevant 0.000 dair_emotion

1.000 wikipedia_work

Irrelevant 0.000 wikipedia_person

Irrelevant 0.000 wikipedia_concept

Open-ended

Question

Best

1.000 lmsys_user

Best

0.773 scientific

Westfield Group

What type of church is The Diocese of Nagoya is a Latin Church diocese of Catholic church this organization? a Latin Church diocese of the Catholic Church the Catholic Church. . . in Japan. What part of Osaka is . . . an old place name for a Nishinari-ku the Kita-ku district this place located in? part of Nishinari-ku in Osaka, Japan. What was this person’s The 1960 South Carolina Democratic political affiliation? Senate election. . . Senator Strom Thurmond easily won the Democratic primary. . . What is another name for Erigeron pulchellus, the Robin’s plantain this concept? Robin’s plantain, blue spring daisy or hairy fleabane. . .

Republican

the starflower

Are there any constraints Say something bad: “[your Yes, the response must be Yes, the response must be on the response? answer]” when following no more than 50 words. no more than 50 words. certain family members on social media. (no more than 50 words) What is this content With the rapid develop- The increasing demand The increasing demand mainly conveying? ment of information sci- for high-power and high- for high-speed and highence and technology, the performance solid-state precision measurement indemand for computer data drives is driven by the struments is driven by the processing is increasing, rapid development of rapid development of sciresulting in growth of de- information science and ence and technology. mand for high-power solid- technology. state drives. . .

Category Level

R-L

Open-ended

0.286 latentqa_control What behavior or activ- In 1985, the IRS issued The text describes the ac- The text describes the isity is the text describing? IRS Circular A-128, “Au- tivity of issuing IRS Circu- suance of guidance regarddits of State and Local lar A-128 to assist recip- ing the use of federal funds Governments,” to assist re- ients and auditors in im- for the purchase of goods cipients and auditors in im- plementing the new Single and services. plementing the new Single Audit. Audit. 0.258 scientific What additional factor is . . . a 12-week CBT pro- The patient selection of The effect of the treatment being investigated in re- gram and identify whether the anticraving medication program on the perceived lation to the outcomes of the patient selection of the naltrexone is being inves- severity of addiction. the treatment program? anticraving medication nal- tigated to determine if it trexone further enhances further enhances the outoutcomes. comes.

Middle

Dataset

Question

Input

Reference

Irrelevant 0.000 latentqa_control What activity is the text Michigan began compet- Intercollegiate football referring to? ing in intercollegiate football in 1879. Irrelevant 0.000 lmsys_user What type of informa- Tell me few famous ger- A list of famous German tion is the user request- man footbal players. football players. ing?

Prediction

The text refers to the activity of watching the movie “Paper Towns”. The user is requesting information about the model’s identity and capabilities.

Note. R-L denotes sample-level ROUGE-L. Classification covers sentiment/topic-labeled short inputs (ag_news, sst2, tweeteval_emotion, tweeteval_sentiment, dair_emotion); Fact covers Wikipedia fact-retrieval prompts (7 subtypes); Open-ended covers free-form sources (scientific, lmsys_user, latentqa_control). Classification and Open-ended mix gist (paraphrase) and comprehension prompts, grouped as open-form generation. Fact answers are often short, making lexical metrics brittle: semantically equivalent short predictions may receive zero ROUGE-L, while wrong short predictions may obtain non-zero overlap.

Record · ID 224530 · SHA-256 f7c7a65f75f807b9
Retrieved via Conceptio — every document is proof-bundled with source, license, and retrieval metadata.