1
PERA: A Perceive-Reason-Act Interface Bridging Sensing, Cognitive Reasoning, and Trustworthy Agentic Response for 6G
arXiv:2607.16896v1 [cs.NI] 18 Jul 2026
Mohammad Farzanullah, Melike Erol-Kantarci, Fellow, IEEE and Lajos Hanzo Life Fellow Abstract—The realization of next-generation (NG) networks hinges on a fundamental departure from pre-programmed protocol engineering towards a paradigm of self-consciously evolving, autonomous and trusted intelligence. While conventional machine learning (ML) has introduced localized automation, it remains inherently bounded by single-task processing pipelines incapable of handling complex cross-layer dynamics. As a partial remedy, large language models (LLMs) excel at generalized cognitive reasoning, but to a degree they remain detached from the rich modalities of wireless telemetry. As a solution, we unveil Generative Network Intelligence conceptualized via the Perceive-Reason-Act (PERA) paradigm. This paradigm treats the wireless channel and the underlying network states as a continuous, multimodal narrative. By synchronizing the perceptual grounding of Large Wireless AI Models (LWAMs) with the cognitive reasoning of LLMs, PERA heralds the era of native NG intelligence. Crucially, this unified intelligence replaces fragmented, task-specific edge models by an efficient multi-task architecture delivering the real-time control needed for supporting dynamic physical applications while reducing both the complexity and energy dissipation. Moreover, we contrast the structural limitations of traditional ML to generative paradigms, conceive agentic reasoning across a NG protocol stack, and detail a practical three-tier design specifically engineered for the resource-constrained wireless edge. This architectural paradigm serves as a foundational framework for realizing fully autonomous, embodied agentic AI in NG networks. To validate this vision, our case study evaluates linkstate classification and beam prediction, demonstrating how grounding wireless telemetry within a cognitive engine delivers the transparent, human-readable rationales required for trusted physical-layer diagnostics and beam control.
I. Introduction The 6th Generation (6G) network is not merely an advanced version of 5th Generation (5G). It represents a fundamental shift from a deterministic bit-pipe infrastructure to fully autonomous, Artificial Intelligence (AI)-native, agentic system. As shown in Fig. 1a, 6G deployments must simultaneously support massive device connectivity, ultra-reliable low-latency communications, and real-time adaptive intelligence across terrestrial and non-terrestrial domains. Conventional networks cannot meet these requirements [1]. Hence 6G must embrace Embodied AI by grounding its cognitive layer within its physical, spectral, and spatial surroundings [1]. This enables agentic nodes to self-consciously perceive, reason, and act (the foundational operations of the PerceiveReason-Act (PERA) paradigm) across the entire layered Mohammad Farzanullah, and Melike Erol-Kantarci are with the School of Electrical Engineering and Computer Science, University of Ottawa, Ottawa, ON K1N 6N5, Canada (e-mail: [email protected]; [email protected]). Lajos Hanzo is with University of Southampton, U.K (email: [email protected]).
protocol stack, as depicted in Fig. 1b, rather than running isolated algorithms in an environmental vacuum. Consequently, the underlying network fabric must evolve to seamlessly synthesize real-time physical-layer variations with high-level cognitive policies across the entire protocol stack. Yet, realizing this at scale requires intelligence at the mobile edge, where latency constraints, communication and computing resource limitations, and real-time control requirements force cognitive reasoning to be grounded in physical-layer realities. Realizing this level of autonomy requires a complete departure from the current status quo of telecommunications AI. Today’s discriminative and reinforcement learning pipelines lack the agentic capability of handling distributed AI tokens or of reasoning over abstract operator intents [2], [3]. Instead, they operate as isolated, task-specific black boxes that collapse under the crosslayer dynamics and high dimensionality of large-scale 6G deployments. Without a bespoke mechanism to orchestrate these complex networking capabilities, conventional frameworks fail to connect high-level intentional reasoning with the continuous, high-dimensional telemetry of the wireless medium. Effectively navigating these network inter-dependencies requires the complementary pairing of two distinct foundation model families, yet an inherent architectural barrier divides them. Large Language Models (LLMs) excel at abstract intent decomposition, logical deduction, and multi-step planning [4]. However, they remain intrinsically blind to the physical world — unable to natively process continuous wireless features like raw channel state information (CSI) or Radio Frequency (RF) interference patterns. Conversely, Large Wireless AI Models (LWAMs) act as powerful sensory systems capable of tokenizing raw telemetry and capturing physical invariants [1], [5], [6]. However, they lack cognitive reasoning, causal interpretation capabilities, and the capacity to interpret highlevel operator mandates. Because existing framework lacks a principled, synchronized interface between these two domains, the LLM cannot read the continuous wireless medium, and the LWAM cannot actuate goal-directed policies, leaving a critical integration puzzle. Motivated by the desire to bridge this dual-model divide, we introduce the Perceive-Reason-Act (PERA) paradigm. At its core, PERA addresses a fundamental modality mismatch: LLMs excel at semantic reasoning over text, while LWAMs excel at feature extraction from raw physical signals. Rather than forcing one of the models to adapt to the other’s domain, PERA establishes a lightweight semantic bridge — a trainable projector
2
Unified Intelligence and Sensing Control
Multi-step planning
Connecting LWAM and LLM for the entire network stack
Logical Deduction
Application Network MAC
Physical
(a) 6G-enabled business and application verticals
SENSORY BACKBONE (LWAM)
SYNCHRONIZATION BRIDGE (MULTI-MODAL INSTRUCTION TUNING)
COGNITIVE CORE Intent (LLM) Decomposition
ISAC perception backbone
Semantic Intelligence
Generative Capability
Mission Oriented Networks
Intent-based networking
Self-healing orchestration
Multi-agent negotiation
Smart packet routing
Massive MIMO coordination
Hierarchial controllers
Sensingassisted scheduling
AI-driven access control
Recongurable Intelligent Surfaces
Generative channel estimation
Dynamic spectrum sharing
Intelligent RF wave beamforming understanding
ISAC perception
Agentic Task Composition
Knowledge Mapping Predictive mobility management
(b) Synchronization of cognitive reasoning and sensory perception across the protocol stack.
Fig. 1. Framework for Embodied AI-driven 6G intelligence: (a) application domains and vertical deployments, and (b) synchronization of cognitive reasoning and sensory perception across the full network stack.
that maps high-dimensional physical features into the embedding space of text tokens. To the LLM, these “network pseudo-tokens” appear as an ordinary sequence of words describing the state of the environment. This alignment transforms PERA into a unified reasoning engine capable of ingesting network pseudo-tokens at the mobile edge, synthesizing cross-layer network state, and producing actionable outputs grounded in both physical observation and semantic reasoning, thereby enabling embodied, intelligent decision-making across the 6G protocol stack. Consolidating this generative interface at the edge yields compelling operational advantages over fragmented, legacy automation. Aligning wireless features with natural language representation can allow network operators to apply advanced prompt engineering—like in-context learning (ICL) and chain of thought (CoT) reasoning [4]—directly over models trained on raw wireless data. This eliminates hosting hundreds of task-specific diagnostic heads, drastically reducing edge memory bottlenecks and overhead. Instead, a single foundational core adapts to novel scenarios and anomalies through contextual, natural-language instructions alone, completely bypassing continuous gradient updates or immediate retraining. The proposed PERA architecture enables these capabilities through unified perception-reasoning integration. Furthermore, by appending recent history to the prompt, the architecture leverages its attention mechanism for realtime predictive decisions while natively delivering humanreadable, causal explanations that demystify black-box network decisions. The remainder of this article provides a comprehensive framework for fully autonomous reasoning and control across the 6G stack. We first establish a structured, layer-wise categorization of the communication stack, contrasting traditional machine learning limitations against
generative paradigms across the physical, network, and application layers. Following this, we detail the practical three-tier design of the PERA architecture, detailing the mechanics of the universal perceptual encoder, the lightweight modality bridge, and the edge-optimized cognitive engine. Finally, we present an empirical case study on beam prediction and line of sight (LoS) classification using the DeepMIMO dataset to demonstrate how grounding wireless signals within a language framework enables zero-shot environmental adaptability. Furthermore, it delivers the transparent, human-readable AI explanations necessary to successfully audit and troubleshoot channel decisions, validating the embodied reasoning paradigm at the core of PERA. II. A Holistic Investigation of the 6G Stack: Synergizing LLMs and Large Wireless AI Models Although recent surveys, most notably [1], establish a comprehensive taxonomy for AI-native 6G, they focus heavily on adapting LLMs to wireless tasks via prompting. This approach often overlooks signal-native models, such as the Large Wireless Model (LWM) [6], which are pretrained directly on physical layer representations such as raw CSI or baseband spectrograms. Moreover current literature relies on isolated automation pipelines that lack the capacity to link high-level reasoning with deterministic physical-layer control. This targeted review identifies these integration gaps to justify the necessity of the unified, edge-deployed architecture proposed in this work. A summary of the works across the network stack is provided in Table I. A. Physical Layer: Foundation Models and Channel Representation In 6G, the Physical (PHY) layer is envisioned to shift from rigid mathematical models toward universal, genera-
3
tive foundation models that leverage task-agnostic feature extraction for rapid adaptation and out-of-distribution robustness. As foundational pillars, the LWM [6] employs Masked Channel Modeling to extract universal wireless features, while the multimodal Wireless Multimodal Foundation Model (WMFM) [5] uses contrastive learning to align radio channels with visual imagery for Integrated Sensing and Communication (ISAC). By applying Perimeter Efficient Fine-Tuning (PEFT) strategies—freezing the backbone and updating only lightweight heads—both models minimize training overhead while outperforming traditional benchmarks in localization and LoS/non-line of sight (NLoS) classification. To eliminate the preprocessing information loss associated with CSI or spectrograms, IQFM [7] operates directly on raw multi-antenna IQ waveforms, capturing spatio-temporal signal characteristics via contrastive selfsupervised learning. Ultimately, transitioning to these compact, task-agnostic encoders is fundamental to mobile edge intelligence, enabling versatile, real-time physicallayer adaptation on resource-constrained devices. B. Network Layer: Intent-Based Multi-Agent Orchestration The Network layer is evolving toward decentralized cognition via multi-agent systems driven by sophisticated orchestration frameworks. This paradigm shift replaces static, imperative configurations with a goal-oriented operations, enabling the network to autonomously decompose complex operator intents into coordinated actions across distributed edge nodes. • Intent-Driven Management: A critical advancement in this layer is the utilization of state-space models for intent-to-policy mapping. As detailed in [2], the linear scaling and efficient long-range dependency handling of Mamba-based architectures allow the network to interpret abstract human intents into realtime configuration scripts without the computational overhead of standard transformers. This generative approach ensures that the network agent can maintain state awareness across dynamic Radio Access Network (RAN) conditions. • Collaborative AI-native Orchestration: Extending this cognition across the cloud-edge continuum requires heterogeneous LLMs to operate in coordination, where lightweight edge models enrich user prompts with location-specific context before forwarding them to a more capable cloud model for full response generation [3]. In this framework, a dedicated Computing Plane jointly schedules communication and computation across the hierarchy, balancing the load between edge and cloud based on task complexity rather than replicating full model copies at every node. This hierarchical synergy yields intelligent network management as a byproduct, with edge LLMs continuously inferring operator intents and predicting traffic demand without any additional orchestration overhead.
•
Intent-Driven Dynamic Network Slicing: To manage 6G complexity, the framework in [8] adaptively segments physical resources across the RAN, transport, and core. It utilizes customized LLMs to iteratively translate operator intents into deployment specifications. Backed by data-driven capacity predictions, this collaborative dialogue allows users to dynamically refine slicing requests, balancing service satisfaction with resource efficiency.
C. Application Layer: Generative Reliability and Semantic Intelligence The Application layer in AI-native 6G systems marks a fundamental shift toward Generative Network Interfaces, where the network state is treated as a multimodal narrative rather than a set of isolated bit-streams. This transition is anchored by generative diffusion models, which provide a robust framework for network state reconstruction and resilience against extreme throughput variations [9]. By treating signal recovery as an inverse probabilistic process, these models can reconstruct highfidelity telemetry even from heavily compressed semantic payloads, ensuring generative reliability for missioncritical edge applications. Complementing this transition is the integration of edge-based large AI models that facilitate cognitive multimodal semantic communication [10]. These frameworks utilize multimodal tokenizers to eliminate bit redundancy, transmitting only the core semantic meaning of a payload to surpass classical Shannon limits. However, the nondeterministic nature of these agentic systems necessitates rigorous verification. To address this, the 6G-Bench framework provides an open, standardized benchmark for evaluating semantic communication and network-level reasoning [11]. By assessing foundation models (FMs) across dozens of complex network reasoning tasks, 6GBench validates the integrity of agentic intent parsing and multi-agent coordination, providing the empirical rigor required to transition from discriminative AI to reasoningcentric 6G orchestration.
D. Prompt Engineering techniques in the wireless domain Integrating LLMs into wireless networks via prompt engineering—such as ICL and CoT—facilitates resourceefficient management by bypassing the computational demands of fine-tuning [4]. At the physical layer, the RF-GPT framework maps IQ waveforms to visual spectrograms, identifying technologies like 5G and LTE with over 97% accuracy while providing grounded explanations [12]. Similarly, applying CoT reasoning to traffic prediction enhances forecast stability by decomposing temporal dependencies, outperforming standard baselines by up to 22.41% in R2 -score [13]. These developments signal a shift toward unified 6G interfaces where LLMs natively query and explain complex spectral dynamics in real-time.
4
TABLE I Layer-wise Literature Categorization: Features, Challenges, and Applications Category
Main Features
Challenges
Applications
Physical Layer Foundation Models [5]–[7]
•
Task-agnostic transformer encoders using Masked Channel Modeling and contrastive selfsupervised learning. • Multimodal alignment of radio channel coefficients with visual imagery for ISAC. • Direct operation on raw IQ signals, bypassing preprocessed inputs such as CSI.
•
High computational overhead for real-time edge inference. • Generalization across heterogeneous propagation environments with minimal retraining. • Capturing joint temporal and spatial characteristics from raw waveforms.
•
Intent-Based Network Orchestration [2], [3], [8]
•
Mamba-based state-space models for efficient, lowoverhead intent-to-policy mapping. • Heterogeneous LLMs distributed across the cloudedge continuum with a dedicated Computing Plane. • Iterative negotiation with residual capacity prediction for end-to-end network slicing.
•
Maintaining transactional integrity and safe rollbacks under dynamic RAN conditions. • Coordination of multi-agent negotiations across RAN, transport, and core domains. • Balancing edge-side prompt enhancement with cloud-side generation under resource constraints.
•
Generative and Semantic Application Layer [9]–[11]
•
Generative diffusion models treating signal recovery as an inverse probabilistic process. • Multimodal tokenizers to eliminate bit redundancy and transmit only core semantic meaning. • Open standardized benchmark evaluating foundation models across complex network reasoning tasks.
•
Reconstructing high-fidelity telemetry from heavily compressed semantic payloads. • Non-deterministic agentic behavior requiring rigorous empirical verification. • Defining evaluation metrics for multi-agent coordination and agentic intent parsing.
•
Prompt Engineering for Wireless Systems [4], [12], [13]
•
ICL, CoT, and self-refinement strategies enabling complex optimization via forward passes only. • Multimodal grounding mapping IQ waveforms to timefrequency spectrograms via STFT. • Plan-based CoT pipeline generating intermediate rationales for temporal dependency decomposition.
•
Avoiding fine-tuning overhead on power-limited and resource-constrained edge devices. • Handling spectral diversity across heterogeneous wireless technologies. • Ensuring forecast stability across non-stationary mobile traffic patterns.
•
E. Synergetic Empowerment: The Unified Organism
These trajectories converge into Synergetic Empowerment, transforming the 6G network into a unified architecture where LLMs provide the cognition and LWAM serve as the sensory body. Fully realizing this embodiment requires a seamless bridge between high-level reasoning and wireless actuation. The following section proposes a framework where instruction tuning acts as the critical connective tissue, utilizing prompt-based AI to align agentic intent with the multi-modal complexities of the wireless medium.
LoS/NLoS classification and channel-based localization. • Zero-shot multi-task sensing without manual labels. • Task-agnostic embeddings for versatile physical-layer edge intelligence.
Converting abstract operator intents into real-time RAN configuration scripts. • Self-healing and autonomous network orchestration. • Dynamic end-to-end network slicing with user-collaborative refinement.
Resilience against extreme throughput variations for mission-critical edge applications. • Surpassing classical Shannon limits via cognitive semantic communication. • Validation of agentic intent parsing and multi-agent coordination performance.
Network optimization and prediction without model retraining. • RF technology identification (5G NR, LTE, Bluetooth) exceeding 97% accuracy [12]. • Mobile traffic forecasting with up to 22.41% R2 improvement over ICL baselines [13].
III. The PERA Architecture: Perceive-Reason-Act for Edge Intelligence Realizing Embodied AI for Mobile Edge General Intelligence requires moving beyond fragmented, single-task models. The diversity of mobile networks—from dynamic physical layer signal variations to application layer traffic orchestration—demands a unified framework. We therefore propose the PERA architecture. PERA bridges raw, multi-dimensional network states across any protocol stack layer with the reasoning capabilities of LLMs, replacing discrete task-specific architectures with a generalized, cross-layer framework explicitly designed for the memory, computational, and latency constraints of heterogeneous Mobile Edge General Intelligence
5
Human Text Prompts (e.g., System Prompt, User Instruction, Operational Requirements)
Raw Network Telemetry (e.g., CSI, continuous wave I/Q, Network layer traffic conditions)
Control Parameters (e.g., routing paths, beamforming
Layer 1: Universal Perceptual Encoder
Layer 2: Learnable Bridge
(e.g., task-agnostic foundation
to share LLM dimensional space)
(Learnable MLP, project features
model, LWAM sensory backbone)
[System Prompt] + [Network Pseudo Tokens] + [User Instruction]
Layer 3: Cognitive Reasoning Core
(Casaule Rationale,
joint reasoning, generates
Troubleshooting)
explanations)
Train the Layer 2
MEC Server: -Host for Layer 3: Cognitive Reasoning Core. -Uses PEFT for finetuning LLM for faster training
Natural Language Explanations
(Pretrained LLM + LoRA, performs
Deployment
Edge Device: -Hosts Layer 1: Universal Perceptual Encoder - reduces overhead of raw data.
indices, modulation coding)
Generative Response
Stage 1: Feature Alignment (State captioning, learn basic vocabulary, freeze encoder/LLM)
Example
Train the Layer 2 and LoRA adapters
Stage 2: Goaloriented Finetuning
Analyze current network topology tokens and historical traffic. Reason about the current bottleneck and predict the optimal routing path to minimize latency for a mission-critical URLLC payload.
(Fine-tune on diverse goal-directed tasks)
1
2
Q
3
W A
123
4
E
5
R
6
T
7
Y
8
U
9
I
D
F
G
H
J
K
Z
X
C
V
B
N
M
Reasoning: "Based on congestion signatures in pseudo-tokens, Path A has high jitter due to a rogue node."
0
O
S
Actuation (Control Parameter): "Optimal decision is Path C, bypassing the congested gateway."
P L
.
Multi-Modal Instruction Tuning
Fig. 2. Overview of the PERA framework’s unification strategy, depicting the two-stage training protocol: Stage 1 (Feature Alignment) and Stage 2 (Goal-oriented Fine-tuning) to bridge the gap between physical RF signals and semantic reasoning.
environments. A. The High-Level Philosophy: From Discriminative AI to Generative Network Interfaces The PERA architecture represents a fundamental shift from discriminative AI toward a “Generative Network Interface.” Traditional AI in telecommunications relies on discriminative models that act as opaque black boxes, outputting isolated labels without contextual reasoning. Operating an intelligent network edge therefore requires an Mobile Edge Computing (MEC) server to host hundreds of disparate models, creating memory bottlenecks, orchestration overhead, and poor generalization to novel scenarios. PERA replaces this monolithic approach with a generative paradigm in which the wireless channel and network state are treated as a Multimodal Narrative. An LLM serves as the “Cognitive Core,” supported by specialized foundation models as “Perceptual Sensory Systems.” Rather than emitting context-less labels, the architecture maps raw network state into a semantic latent space, conditioning the LLM to output a joint probability distribution alongside a natural language reasoning chain. A single FM at the edge can thus ingest raw data from any protocol stack layer, reason over dynamic network intents, and produce context-aware execution strategies. B. The Three-Tier Architectural Stack The PERA architecture, shown in Fig. 2, utilizes a three-tier modular design to optimize resource efficiency by distributing tasks between the MEC server and edge devices. This separation of actions ensures that heavy cognitive tasks rely on the MEC server, while light-weight
sensory encoding harnesses edge devices, such as IoT gateways or user equipment. 1) Layer 1: The Universal Perceptual Encoder: The foundational layer consists of domain-specific, yet taskagnostic, Perceptual Encoders. Function: The primary role of this layer is to ingest raw, high-dimensional, and often non-linear network data. Because we design the PERA framework to be completely generic, the input data modality is highly flexible. It may involve physical layer I/Q continuous waves, MAC layer scheduling matrices, or network-layer traffic flow topologies. Since Layer 1 remains frozen at the device, deployment on resource-constrained hardware is achieved through standard model-based compression techniques (quantization, distillation, pruning), making LWAMs practical for inference. Design: The Perceptual Encoder is typically realized through a Transformer-based architecture that utilizes spatial-temporal patching strategies. Through large-scale self-supervised pre-training, this encoder learns to tokenize the raw input, capturing deep physical invariants and structural correlations while ignoring localized noise. Output: The encoder produces a sequence of highdimensional latent vectors. These vectors represent the underlying structural features of the network environment abstracted entirely from the raw telemetry. 2) Layer 2: The Modality Bridge (Semantic Projector): Function: The fundamental challenge is an inherent language barrier: LLMs “speak” the language of human semantics (text), whereas the Perceptual Encoder “speaks” the language of physics and network dynamics. The Modality Bridge translates between these domains by mapping wireless features into the LLM’s text embedding space.
6
Mechanism: This layer is implemented via a lightweight adapter having significantly fewer parameters than the LLM and Perceptual Encoder. This efficiency makes it computationally suitable for edge deployment without sacrificing alignment quality. Role: It projects the perceptual vectors into a sequence of “Network Pseudo-Tokens.” These pseudo-tokens are mapped to share the exact dimensional space as the LLM’s standard text embeddings. Consequently, to the LLM, the raw network telemetry simply appears as a sequence of words describing the state of the environment in an unfamiliar, yet mathematically justified text. 3) Layer 3: The Edge-Optimized Cognitive Reasoning Core: Function: The apex of the PERA architecture is the Cognitive Reasoning Core, instantiated by a pre-trained LLM. Its objective is to perform joint reasoning over human-readable text instructions and the network pseudotokens generated by Layer 2. Input: The LLM ingests a concatenated sequence of multimodal inputs structured as: [System Prompt] + [Network Pseudo-Tokens] + [User/System Instruction]. The instructions can be provided by operators, derived from network policies, or generated by LLM agents. Output: The model outputs a comprehensive generative response that includes both deterministic control parameters (e.g., precise routing paths, specific modulation and coding schemes, and beamforming indices) alongside natural language explanations detailing its logical reasoning process. To ensure that this architecture is feasible within the strict memory limits of a Mobile Edge node, the Cognitive Core leverages PEFT. Specifically, the foundational weights of the LLM remain entirely frozen. Adaptation to specific network tasks is achieved using Low-Rank Adaptation (LoRA) modules injected into the attention layers. By updating only a minuscule fraction of the model’s total parameters, PERA enables dynamic, real-time task switching; the edge server can swap out lightweight LoRA adapters in milliseconds without needing to unload and reload massive foundation models in its limited memory. C. The Unification Strategy: Multi-Modal Instruction Tuning (MM-IT) The innovation of the PERA architecture lies not just in its individual structural components, but in the methodology used to fuse them. In order to bind the frozen perceptual encoder, the lightweight Modality Bridge, and the LoRA-adapted LLM core, we use MM-IT adopted from [14]. This two-stage protocol is designed to force the LLM to comprehend raw network telemetry as a foundational context for complex problem-solving. Unlike monolithic instruction tuning, MM-IT’s separation of semantic alignment and task-specific reasoning enables explicit grounding in physical reality, essential for network safety and reliability.
1) Stage 1: Representation Alignment: Before an LLM can effectively optimize a network, it must first inherently understand what the mathematical representations of network data actually mean. The goal of Stage 1 is to teach the LLM the fundamental vocabulary of the telecommunications stack. Task: During this phase, the system performs a task analogous to “State Captioning.” A massive dataset of paired network states and text descriptions is utilized. For example, given a latent representation of a complex multipath fading environment or a highly congested routing table, the target output is a precise descriptive text (e.g., “This sequence represents a high-mobility, non-lineof-sight signal path with severe Doppler shifting”). Mechanism: Only the Modality Bridge (Layer 2) is trained during this stage. The Perceptual Encoder and the LLM remain entirely frozen. The loss function applied is the standard Causal Language Modeling loss over the predicted description tokens. This phase acts as a translation dictionary, mechanically mapping the geometric properties of the network latent space into the semantic topology of the LLM’s word embeddings. 2) Stage 2: Goal-Oriented Instruction Tuning: Once the semantic alignment is established, the system must learn to execute complex, goal-directed edge-intelligence tasks. Goal: To solve specific networking, communication, and orchestration challenges using natural language reasoning and instructions. Mechanism: In this stage, the Modality Bridge remains trainable, and specific LoRA adapters are activated within the LLM. The model is fine-tuned on a diverse dataset of instruction-response pairs covering multiple domains, such as ISAC, predictive handover optimization, and dynamic bandwidth allocation. Instruction Format: A training sample in this phase takes the following format: • Instruction: “Analyze the channel state and determine the optimal beam group.” • Response: “Energy peaks at the center with high L2 norm agreement. Beam Group 5 dominates. Select Beam Group 5 for optimal signal alignment.” Through this extensive Instruction Tuning, the architecture transitions the LLM from a passive text generator into an active, embodied agent capable of parsing complex human intents and executing highly logical, deterministic actions within the mobile edge environment. D. Key Advantages for Mobile Edge General Intelligence Transitioning to the unified, generative PERA architecture optimizes the wireless edge by consolidating diverse operations—such as localization and congestion control— into a single foundational model, drastically reducing edge server memory and overhead. Driven by LLM reasoning, its zero-shot adaptability allows the network to fluidly handle novel environments and anomalies through contextual instructions without requiring immediate retraining. Furthermore, PERA solves the telecom “black-box”
7
IV. Case Study: PERA Implementation via Physical Layer Instruction Tuning To validate the theoretical framework established in the preceding sections, this section presents a comprehensive empirical evaluation of the PERA architecture. We evaluate the performance of this unified architecture within a simulated environment, focusing on the efficacy of the two-stage MM-IT protocol in translating highdimensional network telemetry into actionable, reasoningcentric responses. A. Experimental Setup and Dataset Curation The experimental setup utilizes spatial channel characteristics and LoS/NLoS metadata derived from the DeepMIMO dataset [15]. We employ the LWM [6] as our Universal Perceptual Encoder to ingest complex CSI and high-dimensional RF waveforms. The LWM functions as a universal feature extractor, tokenizing these raw CSI matrices into structured latent representations that capture deep physical invariants of the wireless environment. To bridge the gap between these physical representations and linguistic reasoning, we constructed an instruction-tuning dataset by mapping the tokenized channels into customcurated supervised dialogue pairs. Each instance bundles a natural-language task instruction with the LWMgenerated observation to produce a response defining the link condition and optimal beam index. B. Implementation and Training Protocol The implementation follows a two-phase strategy designed to establish a shared ”dialect” between the wireless and language domains. In the first phase, Representation Alignment, a lightweight modality bridge projects the wireless global summary token into the high-dimensional embedding space of the LLM. This bridge, implemented as a two-layer Multi-Layer Perceptron (MLP) projector with a hidden width of 1024 and GELU activation, is optimized using the AdamW optimizer with a learning rate of 10−4 and a batch size of 4. During this stage, the LLM backbone remains frozen while the bridge learns to map latent wireless features to qualitative text descriptions. In the second phase, Goal-Oriented Instruction Tuning, the model performs joint reasoning over the instruction prompt and the wireless token. To preserve physical
L S/NL S F1 by Scenari
1.0
CNN MLP n LWM t kens Pr p sed Appr ach
0.8
LoS/NLoS F1
dilemma by providing native explainability, generating human-readable, causal rationales for critical actions like bandwidth reallocation to streamline human troubleshooting. These capabilities can translate to measurable performance improvements in classification accuracy, inference latency, and network throughput. Finally, by appending recent history to the prompt, the model leverages its attention mechanism for in-context predictive tracking, dynamically forecasting near-future states while completely bypassing continuous gradient updates.
0.6 0.4 0.2 0.0
ork
Newy
er
Denv
Losa
s
ngele
on
Houst
nix
Phoe
Miam
i
Fig. 3. Comparison of F1 score for LoS/NLoS classification.
fidelity and prevent generative ”hallucinations”, we incorporated auxiliary pathways to predict link state and coarse beam groupings directly from the embeddings. By grounding the language output in physical labels, these auxiliary tasks prevent the model from generating inaccurate or physically inconsistent explanations. The beam-group head utilizes a residual 1-D convolutional neural network (CNN) stem having a width of 256 and five residual blocks, while the LoS head consists of a two-layer MLP. Physical fidelity is enforced through a combined objective function appropriately weighting language modeling, LoS classification, and beam-grouping by a ratio of 0.3 : 1.0 : 0.5, respectively. For this instruction tuning, the LLM is wrapped with LoRA adapters on the query, key, value, and output projections (rank 16, α = 32). Optimization proceeds through a three-stage curriculum of two epochs each, with AdamW learning rates decaying from 10−3 to 10−4 . This staged approach ensures the interface remains faithful to physical labels while optimizing for the multi-task reasoning capabilities of the LLM. For the LLM in the cognitive reasoning core, we use the Qwen-3 Instruct Model, as its advanced intent decomposition and multi-step planning capabilities allow the network to translate abstract human mandates into deterministic configurations while maintaining the high-fidelity reasoning required for explainable 6G intelligence. During the training process, we utilized five scenarios from the DeepMIMO dataset, including San Diego, Santa Clara, Fortworth, Indianapolis, and Oklahoma. To evaluate generalization, the inference was conducted on six novel scenarios. For the beam prediction task, the model was configured to identify 16 distinct beam groups. To rigorously assess the performance of the PERA architecture, two benchmarks were developed for comparison. The first benchmark employs a standard CNN trained on raw CSI data, serving as a baseline for traditional discriminative approaches. The second benchmark trains identical diagnostic heads directly from the LWM representations to isolate the impact of LWM driven feature extraction. C. Experimental Results The results of our empirical evaluations, illustrated in Fig. 3 and Fig. 4, demonstrate the superior performance of the PERA architecture across diverse novel scenarios. Fig. 3 compares the F1 scores for LoS and NLoS classification,
8
1.0
Beam F1 (Macro)
0.8
Beam Predic(ion F1 by Scenario CNN Seq-CNN on LWM tokens Proposed Approach
0.6
0.4
0.2
0.0
york
New
ver
Den
s
on
ngele
Losa
Houst
nix Phoe
Miam
i
Fig. 4. Comparison of F1 score for beam prediction.
Peak @ 122
Stage 1: Perceptual Input (LWM representation)
Inference example from New York Scenario
Valley @ 0
Stage 2: Cognitive Reasoning Core
Link state
Top-K beam grouping Beam 5 (97.4%)
LoS: 99.37% Beam 6 (2.5%)
Beam 10 (0.1%)
Stage 3: Generative Explanation LoS explanation: The energy distribution shows a sharp peak at the center of the ribbon, with high agreement across patch-level L2 norm values (peak@122 and valley@0; CLS aggregate is 0.0523. The link state is clear-view. Energy peaks at central patches, where L2 norms show strong local agreement in the middle of the ribbon.
Beam explanation: Group 5 dominates with high mass, consistent with a stable patch of elevated energy in the central ribbon near index 67–73, where L2 norm peaks align with a structured pattern: a smooth gradient drop from peak at 69 to valley at 0 and 1248.
Fig. 5. Demonstration of the PERA architecture pipeline during inference in an urban scenario. Stage 1 illustrates the perceptual energy distribution extracted from the channel tokens. Stage 2 shows the cognitive reasoning core classifying the link state as LoS (99.37%) and identifying the optimal beam group (97.4%). Stage 3 presents the generative explanations that ground the network’s decisions in the observed physical invariants.
where the proposed approach consistently outperforms the raw CSI-trained CNN by 7.73% and the MLP (LWM) baseline by 11.77% in the environments considered. Similarly, Fig. 4 highlights the framework’s efficacy in beam prediction, showing a substantial performance gain of 31.17% over the baseline CNN and a 9.90% improvement over the Seq-CNN (LWM). These findings indicate that by intrinsically integrating the cognitive reasoning of the LLM with the perceptual grounding of the LWM, the system achieves more agile environmental adaptability and confident decision-making than traditional discriminative models. Furthermore, the stability of these results across spatially distinct scenarios, such as Houston and Miami, validates the architecture’s adaptability to time-variant propagation conditions in the 6G mobile ecosystem. The end-to-end inference pipeline of the PERA architecture is demonstrated in Fig. 5, using a test sample. In Stage 1, the perceptual input extracted from the channel tokens reveals a specific energy distribution. In Stage 2, the Cognitive Reasoning Core processes these representations
to classify the link state as LoS with 99.37% confidence and identifies Beam 5 as the optimal beam group with 97.4% probability. Finally, Stage 3 generates explanations that ground these decisions in observed physical invariants. For instance, the system provides a causal rationale for the LoS classification by identifying high agreement across patch-level L2 norm values, while the beam selection is justified by the stable patch of elevated energy detected in the central ribbon. This transparent reasoning mechanism enables the realization of Explainable AI, eliminating the “black-box” limitation of traditional models by providing a humanreadable, causal rationale that is critical for streamlining auditing and troubleshooting processes in complex network environments. V. Conclusion A generative network interface architecture realizing ”Agentified Embodied” AI for Mobile Edge General Intelligence was conceived for 6G. By synchronizing an LWAM sensory backbone with an LLM cognitive engine via a two-stage MM-IT protocol, PERA overcomes traditional discriminative AI limitations: task-specificity, black-box opacity, and poor generalization. Our cross-layer survey demonstrates this convergence across the protocol stack: physical-layer foundation models offer task-agnostic sensory grounding, intent-based frameworks enable autonomous, agentic orchestration, and application-layer diffusion and semantic communication systems push past classical Shannon limits. Together, they establish a unified network architecture fusing cognition and perception. Empirical validation on the DeepMIMO dataset confirms that PERA consistently outperforms discriminative baselines in previously unseen scenarios, achieving +9.90% in beam prediction and +7.73% in LoS/NLoS classification F1 scores. Crucially, these gains are paired with humanreadable, causal explanations, directly mitigating the explainability concern impeding operator trust in channel diagnostics and beam control. Acknowledgment The authors gratefully acknowledge Dusit Niyato, from the Nanyang Technological University, Singapore, for his insightful feedback and detailed review of this manuscript. His critical suggestions significantly strengthened the technical depth and clarity of the presentation. References [1] F. Zhu, X. Wang, S. Jiang, X. Li, M. Zhang, Y. Chen, C. Huang, Z. Yang, X. Chen, Z. Zhang et al., “Wireless large AI model: Shaping the AI-native future of 6G and beyond,” arXiv preprint arXiv:2504.14653, 2025. [2] M. A. Habib, M. Elsayed, Y. Ozcan, P. E. Iturria-Rivera, M. Bavand, and M. Erol-Kantarci, “Generative AI for IntentDriven Network Management in 6G RAN: A Case Study on the Mamba Model,” IEEE Wireless Communications, pp. 1–8, 2026. [3] Y. Chen, R. Li, Z. Zhao, C. Peng, J. Wu, E. Hossain, and H. Zhang, “NetGPT: An AI-native network architecture for provisioning beyond personalized generative services,” IEEE Network, vol. 38, no. 6, pp. 404–413, 2024.
9
[4] H. Zhou, C. Hu, D. Yuan, Y. Yuan, D. Wu, X. Chen, H. Tabassum, and X. Liu, “Large Language Models for Wireless Networks: An Overview from the Prompt Engineering Perspective,” IEEE Wireless Communications, vol. 32, no. 4, pp. 98–106, 2025. [5] M. Farzanullah, H. Zhang, A. B. Sediq, A. Afana, and M. ErolKantarci, “Wireless Multimodal Foundation Model (WMFM): Integrating Vision and Communication Modalities for 6G ISAC Systems,” arXiv preprint arXiv:2512.23897, 2025. [6] S. Alikhani, G. Charan, and A. Alkhateeb, “Large wireless model (LWM): A foundation model for wireless channels,” arXiv preprint arXiv:2411.08872, 2024. [7] O. Mashaal and H. Abou-Zeid, “IQFM—A Wireless Foundation Model for I/Q Streams in AI-Native 6G,” IEEE Open Journal of the Communications Society, vol. 7, pp. 1426–1441, 2026. [8] L. Liu, A. Rajkumar, and Y. Jiang, “Advanced end-to-end intent-driven dynamic network slicing: Challenges, solutions, and implementation,” IEEE Communications Standards Magazine, pp. 1–8, 2025. [9] D. Fan, R. Meng, X. Xu, Y. Liu, G. Nan, C. Feng, S. Han, S. Gao, B. Xu, D. Niyato, T. Q. S. Quek, and P. Zhang, “Generative diffusion models for wireless networks: Fundamental, architecture, and state-of-the-art,” IEEE Communications Surveys & Tutorials, vol. 28, pp. 5632–5677, 2026. [10] Y. Sun, S. Guo, X. Qiu, J. Chen, Y. Liu, R. Xin, and F. Qi, “Edge large ai model empowered cognitive multimodal semantic communication system,” in ICC 2025 - IEEE International Conference on Communications, 2025, pp. 3406–3411. [11] M. A. Ferrag, A. Lakas, and M. Debbah, “6g-bench: An open benchmark for semantic communication and network-level reasoning with foundation models in ai-native 6g networks,” IEEE Open Journal of the Communications Society, vol. 7, pp. 3305–3330, 2026. [12] H. Zou, Y. Tian, B. Wang, L. Bariah, S. Lasaulce, C. Huang, and M. Debbah, “RF-GPT: Teaching AI to See the Wireless World,” arXiv preprint arXiv:2602.14833, 2026. [13] M. Ghadaksaz, M. Farzanullah, A. B. Sediq, A. Afana, and M. Erol-Kantarci, “Chain-of-Thought Reasoning Enhances InContext Learning for LLM-Based Mobile Traffic Prediction,” 2026. [Online]. Available: https://arxiv.org/abs/2605.09260 [14] H. Liu, C. Li, Q. Wu, and Y. J. Lee, “Visual instruction tuning,” Advances in neural information processing systems, vol. 36, pp. 34 892–34 916, 2023. [15] A. Alkhateeb, “DeepMIMO: A generic deep learning dataset for millimeter wave and massive MIMO applications,” in Proc. of Information Theory and Applications Workshop (ITA), San Diego, CA, Feb 2019, pp. 1–8.