PaperlayerView exact revision on arXiv

Source-linked AI summary

AGENS: Agentic Link Construction for Environment and Intent Aware 6G Communication

Zhaoyang Li, Shangzhuo Xie, Qianqian Yang

arXiv:2610.34131v1eess.SP

TL;DR

Existing AI communication methods mostly focus on isolated tasks or PHY modules, while AI-native 6G links must jointly account for channel conditions and user requirements. AGENS generates full-link PHY configurations within a modular framework using cross-modal fusion, sequential action generation, and two-stage training. Simulations report adaptable, near-optimal strategies and lower inference latency than search-based heuristic methods, with competitive link-level performance.

  • Problem

    Existing AI communication methods mainly target isolated tasks or PHY modules rather than jointly constructing links for dynamic environments, user intents, and coupled PHY functions.

    Source: §I · p. 1

  • Method

    AGENS fuses CSI with natural-language requirements, uses an LLM and LSTM-based sequential action generation, and applies supervised initialization followed by multi-objective reinforcement learning.

    Source: §I · p. 2

  • Results

    Simulations show that AGENS adapts to diverse requirements and challenging channels, rapidly generating near-optimal configurations; it has lower inference latency than search-based heuristics while maintaining competitive link-level performance.

    Source: §I · p. 2, §VII · p. 13

  • Takeaways & Limitations

    AGENS constructs environment- and intent-aware full-link configurations using predefined PHY-module candidates, maintaining compatibility with modular communication systems.

    Source: §VII · p. 12

Abstract

from arXiv · show

The emergence of diverse intelligent services is driving 6G networks toward an intelligent communication ecosystem with increasingly complex operating scenarios. However, existing AI-based communication methods mainly focus on isolated tasks or individual physical-layer mod- ules, which is insufficient for AI-native 6G systems that must jointly perceive dynamic environments, understand specific user intents, coordinate multiple coupled PHY-layer modules, and remain compatible with existing communication infrastructures. To address this gap, this paper, for the first time, explores auto- matic environment- and intent-aware full-link construction, which aims to generate executable PHY-layer configurations according to channel conditions and service requirements. We propose AGENS, an agentic communication decision-making framework for AI-native 6G link construction. Rather than replacing the existing PHY-layer protocol stack with a black-box end-to-end model, AGENS operates on top of the modular communication framework and jointly optimizes end-to-end PHY-layer configu- rations. Specifically, AGENS encodes CSI and natural-language user instructions, aligns communication states with user intents through cross-modal fusion, and leverages a pretrained LLM for intent-aware reasoning. An LSTM-based sequential action generator further captures dependencies among coupled sub- actions and produces structured full-link strategies. To enable stable training in the high-dimensional discrete action space, AGENS adopts a two-stage learning framework, where heuristic supervised initialization is followed by multi-objective reinforce- ment learning to jointly optimize reliability, throughput, and power consumption.

I. INTRODUCTION

Existing AI communication methods often optimize individual PHY modules, while 6G links must coordinate modules across changing environments and diverse user intents. AGENS constructs full-link configurations within a modular PHY framework, and conventional module-centric configuration can require scenario-specific retuning.

I. INTRODUCTION

  • Most existing AI methods optimize individual communication functions, limiting their ability to coordinate coupled PHY modules for dynamic environments and diverse user requirements.LLMs may coordinate structured decisions without replacing mature PHY modules, but existing approaches do not explicitly construct full links conditioned jointly on CSI and user requirements.Source: §I · pp. 1–2
  • AGENS combines CSI and natural-language requirements to generate executable PHY configurations, using cross-modal fusion, an LSTM-based sequential action generator, and two-stage training.Heuristic exploration supplies expert-like samples for supervised initialization, followed by multi-objective reinforcement learning to optimize BER, throughput, and power consumption.Source: §I · p. 2
  • AGENS uses a digital-twin-inspired link simulator for closed-loop evaluation and policy optimization over modular PHY configurations.The framework jointly models CSI and natural-language intent, then generates linked sub-actions with an LSTM-based state model.Source: §I · p. 2
  • Simulations report that AGENS adapts to diverse user requirements and challenging channels while rapidly generating near-optimal link configurations without time-consuming heuristic search.The proposed training pipeline combines heuristic exploration and supervised initialization with multi-objective reinforcement learning.Source: §I · p. 2

A. Traditional Communication Technologies

  • Conventional PHY configurations rely on rules, lookup tables, or thresholds, and often need scenario-specific retuning as channel conditions or service objectives change.Isolated module design makes coupled effects difficult to capture, limiting globally coordinated full-link optimization.Source: §I · p. 1, A. Traditional Communication… · p. 3

B. AI-Empowered Wireless Communications

AI has improved many wireless tasks, but existing methods generally do not construct full links by jointly coordinating coupled PHY modules for CSI and user requirements. This section reviews that gap and describes a modular PHY link whose configurations and performance are represented through structured actions and link-level metrics.

B. AI-Empowered Wireless Communications

  • Most AI wireless studies address a single task, module, or objective rather than coordinating coupled PHY modules under dynamic environments and diverse user requirements.Existing LLM and wireless foundation-model approaches likewise mainly serve as task-specific predictors, single-module optimizers, or representation learners, leaving joint CSI- and intent-conditioned link construction unaddressed.Source: B. AI-Empowered Wireless Communications · p. 3, C. LLMs Empowered Wireless… · p. 3

C. LLMs Empowered Wireless Communications

  • LLM wireless research includes channel and beam prediction, beamforming, resource allocation, and general-purpose channel representation learning, but does not explicitly address full-link construction across coupled PHY modules.The reviewed methods include task embeddings for multiple beamforming objectives and foundation models aimed at transferable or multi-task representations.Source: C. LLMs Empowered Wireless… · p. 3

A. Modular PHY-Layer Communication System

  • The considered link selects predefined strategies for coding, modulation, precoding, power control, channel estimation, and equalization; their joint selection forms an executable PHY configuration.The configuration is represented as a structured action vector whose entries denote the selected strategy for each module.Source: A. Modular PHY-Layer Communication… · pp. 3–4
  • Link performance is evaluated using BER, effective transmission rate, and power compensation overhead.The effective transmission rate is the ratio of information bits to transmitted modulation bits.Source: A. Modular PHY-Layer Communication… · p. 4

B. AGENS Input–Output Formulation

AGENS maps channel observations and textual instructions to a structured action representing a complete PHY-layer configuration. Joint strategy selection is modeled as a reinforcement-learning decision process that learns a policy to maximize expected reward.

B. AGENS Input–Output Formulation

  • The input spaces are channel observations for CSI and textual instructions describing communication requirements.The input variables are H for CSI and T for the instruction.Source: B. AGENS Input–Output Formulation · p. 4
  • AGENS maps CSI and a textual user instruction to a structured action selected from the candidate strategies of configurable PHY modules.The action is the full-link configuration.Source: B. AGENS Input–Output Formulation · p. 4
  • AGENS formulates the decision process as a multimodal mapping from channel observations and user-intent descriptions to a structured full-link action.The policy πθ parameterizes this mapping.Source: B. AGENS Input–Output Formulation · p. 4

C. Reinforcement Learning-Based Decision Modeling

  • Joint strategy selection across communication modules is modeled as a reinforcement-learning decision process.The process is defined using state and action spaces and a reward function.Source: C. Reinforcement Learning-Based… · p. 4
  • The reward-maximizing objective is to learn an optimal policy mapping states to actions.The objective maximizes expected reward over the state distribution and policy-selected actions.Source: C. Reinforcement Learning-Based… · p. 4
  • The policy assigns each action a probability conditioned on the state.πθ(a|s) is the probability that policy πθ selects action a under state s.Source: C. Reinforcement Learning-Based… · p. 4

IV. PROPOSED METHOD

AGENS combines CSI and text encoders, cross-modal fusion, a pretrained language-model backbone, and an LSTM action generator to produce structured PHY strategies. Its multimodal encoders transform channel observations and user instructions into representations suitable for joint processing.

IV. PROPOSED METHOD

  • AGENS has CSI and text encoders, cross-modal fusion, a pretrained language-model backbone, and an LSTM generator for structured PHY strategies.The components respectively support multimodal perception, alignment of channel states with user intents, intent-aware representation learning, and strategy generation.Source: §IV · p. 4

A. Multimodal Encoders

  • The CSI encoder extracts channel-aware representations, while the text encoder extracts semantic-aware representations of user instructions.The text is tokenized and embedded using the pretrained LLM tokenizer and embedding layer.Source: A. Multimodal Encoders · pp. 4–5
  • A temporal patching operation serializes the CSI sequence to facilitate transformer-style processing.L and F denote the temporal length and feature dimension of CSI observations.Source: A. Multimodal Encoders · p. 4
  • Cross Spatial Self-Attention models spatial-frequency correlations and inter-channel interference patterns in the CSI representation.The resulting Y representation is described as the enhanced CSI feature representation after spatial attention and structural reshaping.Source: A. Multimodal Encoders · p. 4
  • A linear projection aligns the enhanced CSI representation with the LLM semantic embedding space, yielding CSI embeddings of dimension d.The CSI embedding is denoted YCSI ∈ R^L×d.Source: A. Multimodal Encoders · p. 5

B. Cross-Modal Fusion

AGENS compresses and filters textual representations conditioned on CSI, then aligns text semantics with channel features through cross-modal attention. The fused multimodal input is refined by a pretrained Transformer backbone to model dependencies between PHY environments and communication requirements.

B. Cross-Modal Fusion

  • A lightweight connector compresses and filters text representations conditioned on CSI to reduce the overhead and redundant semantics of long instructions.The first step projects textual embeddings to a shorter sequence, producing Ztext.Source: B. Cross-Modal Fusion · p. 5
  • Cross-modal attention uses CSI features as queries and textual embeddings as keys and values to align channel features with text semantics.The attention mechanism is inspired by the Q-Former structure.Source: B. Cross-Modal Fusion · p. 5

C. LLM Backbone for Intent-Aware Reasoning

  • The CSI-aligned text representation is concatenated with CSI embeddings to form the LLM's fused multimodal input.This input is denoted X0.Source: C. LLM Backbone for Intent-Aware… · p. 5
  • The pretrained Transformer backbone progressively refines multimodal representations to capture high-order dependencies between PHY environments and semantic communication requirements.Its layers use multi-head self-attention and feed-forward networks.Source: C. LLM Backbone for Intent-Aware… · p. 5

D. LSTM-Based Sequential Action Generator

AGENS generates a full PHY configuration as a sequence of dependent sub-actions rather than selecting each module's action independently. An LSTM conditions each action on the multimodal state and previously generated actions, avoiding enumeration of the full joint action space.

D. LSTM-Based Sequential Action Generator

  • The LSTM-based generator models dependencies among PHY modules instead of using independent actor heads for sub-actions.Its autoregressive mechanism preserves dependencies while avoiding enumeration of the full joint action space.Source: D. LSTM-Based Sequential Action… · pp. 5–6
  • The overall PHY configuration is a multidimensional discrete action space, with each sub-action corresponding to a strategy for one PHY module.The joint policy is factorized into conditional distributions over the sequence of sub-actions.Source: D. LSTM-Based Sequential Action… · p. 6
  • A compact embedding of the LLM multimodal representation initializes the sequential generation process and the LSTM hidden state.The initial state uses a learnable linear projection.Source: D. LSTM-Based Sequential Action… · p. 6
  • At each step, the LSTM uses its previous hidden state and the prior action embedding to produce a categorical distribution for the next action dimension.The selected sub-action is fed into the next decoding step, and all generated sub-actions form the complete PHY link.Source: D. LSTM-Based Sequential Action… · p. 6

V. TRAINING METHOD

AGENS treats full-link configuration as a one-step reinforcement-learning decision conditioned on CSI and a user instruction. It generates a structured sequence of PHY-module choices and scores the resulting link using intent-weighted reliability, throughput, and energy-efficiency metrics.

A. Reinforcement Learning Formulation

  • AGENS formulates full-link configuration as a one-step decision in which the agent jointly configures PHY modules from the communication environment and user requirement.The sampled instance index is not a multi-step temporal transition.Source: A. Reinforcement Learning Formulation · p. 6
  • The state combines CSI Ht with user instruction Tt, while the action is a complete configuration assembled from module-specific sub-actions.The policy generates sub-actions sequentially, conditioning each choice on previously generated sub-actions.Source: A. Reinforcement Learning Formulation · p. 6
  • The simulator evaluates each configuration using BER, rate, and power overhead, which are combined into a normalized weighted reward.The reliability, throughput, and energy-efficiency weights sum to one and are selected according to user intent.Source: A. Reinforcement Learning Formulation · pp. 6–7
  • Because the decision is single-step, policy learning maximizes expected reward over the state distribution induced by sampled CSI and user-intent prompts.The objective is stated as maximizing expected reward under the policy.Source: A. Reinforcement Learning Formulation · p. 7

B. Two-Stage Training

AGENS is trained first to imitate heuristic expert configurations and then fine-tuned with PPO. The heuristic generates sequential expert sub-actions, while PPO uses an entropy estimate that avoids enumerating the full joint action space.

B. Two-Stage Training

  • The two-stage training framework first uses heuristic samples for supervised fine-tuning, then applies PPO to improve the initialized policy.PPO includes entropy regularization.Source: B. Two-Stage Training · p. 7
  • The greedy heuristic builds expert-like configurations by selecting module sub-actions sequentially to locally maximize communication utility.Each candidate sub-action is evaluated in a complete action sequence; proceeding across dimensions avoids full joint-action enumeration and reduces search complexity.Source: B. Two-Stage Training · p. 7
  • Teacher-forcing supervised fine-tuning trains each sub-action predictor on expert action prefixes, initializing the policy for reinforcement learning.Expert-like state-action tuples are stored in a replay buffer and used for imitation.Source: B. Two-Stage Training · p. 7
  • PPO fine-tuning uses the immediate reward as the advantage in this single-step decision process and clips policy updates relative to the previous policy.The clipping factor limits deviation from the previous training iteration's policy.Source: B. Two-Stage Training · p. 7
  • A sampled conditional-entropy estimate encourages exploration without enumerating all joint actions.It requires only the per-step categorical distributions produced by the policy network.Source: B. Two-Stage Training · pp. 7–8

VI. NUMERICAL RESULTS

The numerical-results section evaluates AGENS through baseline comparisons, ablations, and a human–machine interaction example. It also identifies tables of CSI-generation parameters and PHY-layer configuration options.

VI. NUMERICAL RESULTS

  • The evaluation is organized around baseline comparisons, ablation studies, and a human–machine interaction example.These evaluations are intended to assess effectiveness, adaptability, and flexibility.Source: §VI · p. 8
  • Table I summarizes the simulation parameters used to generate CSI datasets.Source: Table I · p. 8
  • Table II lists candidate configuration options for PHY-layer modules.Source: Table II · p. 8

A. Experimental Setup

The experiments use CSI and user-intent samples across three urban scenarios, then compare strategies under throughput, reliability, and energy-saving preferences. AGENS shows intent-specific trade-offs and coordinated performance, while the Transformer-based variant is less adaptive.

A. Experimental Setup

  • CSI datasets cover Munich, Etoile, and Florence, with 12,000 training, 2,000 validation, and 4,000 test samples.CSI is generated with Sionna-RT across heterogeneous channel and geometric propagation conditions, with user-intent descriptions assigned to samples.Source: A. Experimental Setup · p. 8
  • The evaluation covers high-throughput, high-reliability, and energy-efficient preferences, and uses GPT-2 Medium with 355M parameters as the LLM backbone.The paper states that the framework is not tied to a specific LLM and can use other LLMs.Source: A. Experimental Setup · p. 8

B. Model Decision Performance

  • Under high throughput, AGENS achieves the highest rate across most SNR regimes, especially medium and high SNR, while keeping BER within an acceptable range.The comparison includes the strategy-selection methods shown in Fig. 4.Source: Figure 4 · p. 9, B. Model Decision Performance · pp. 9–10
  • Under high reliability, AGENS attains nearly zero BER across the SNR range with more conservative configurations, at the cost of lower transmission rate.Fig. 5 compares strategy-selection methods for this intent.Source: Figure 5 · p. 9, B. Model Decision Performance · p. 10
  • For energy saving, AGENS selects lower power compensation and more conservative modulation and coding, balancing power overhead, BER, and achievable rate.The comparison of strategy-selection methods for this intent is shown in Fig. 6.Source: Figure 6 · p. 9, B. Model Decision Performance · p. 10
  • The Transformer-based variant performs worse than heuristic approaches under throughput or reliability priorities and makes rigid, nearly identical rate choices across scenarios under energy saving.Fig. 6 shows this variant's energy-saving behavior; the paper reports that its strategies can produce excessively high BER.Source: Figure 6 · p. 9, B. Model Decision Performance · p. 10
  • AGENS produces smoother metric curves than random and greedy strategies, which can fluctuate because of randomness and local module-level optimization.The paper associates the smoother curves with better global coordination and higher decision stability.Source: B. Model Decision Performance · p. 10
  • Across the three intents, AGENS produces distinct performance trends, matching reliability, throughput, and energy-saving preferences with different trade-offs.The reported strategies respectively prioritize lower BER, higher rate despite increased BER and power use, or reduced power compensation with moderate rate and reliability.Source: B. Model Decision Performance · p. 10

C. Adaptation to Heterogeneous Channel Scenarios

AGENS infers a throughput-oriented intent from a natural-language prompt and combines it with CSI and SNR to generate PHY strategies. Across a benign near-static and a challenging mobile rich-scattering scenario, it adapts selected modules while retaining the requested intent.

C. Adaptation to Heterogeneous Channel Scenarios

  • For throughput-prioritization prompts, AGENS jointly considers CSI and SNR, inferring the intent from natural-language text rather than a directly supplied preference label.Scenario A uses SNR = 5 dB, while Scenario B uses SNR = −5 dB.Source: C. Adaptation to Heterogeneous Channel… · p. 10
  • The evaluation uses a near-static LoS-like scenario and a long-range, mobile, rich-scattering NLoS scenario.Scenario A uses distances of 10–60 m and maximum multipath depth 1; Scenario B uses 200–450 m and depth 8.Source: C. Adaptation to Heterogeneous Channel… · p. 10
  • AGENS changes modulation, coding rate, channel estimation, and precoding between scenarios while preserving throughput-oriented intent.Other components remain unchanged; the comparison demonstrates adaptation to heterogeneous propagation conditions.Source: C. Adaptation to Heterogeneous Channel… · p. 10, Table III · p. 11

D. Generalization to Unseen-City Scenarios

Across the reported scenarios, AGENS generates competitive link configurations, adapts strategies to user preferences, and runs faster than search-based heuristics. The section also reports ablation, inference-time, and interaction results, alongside the framework’s claimed compatibility and adaptability.

E. Ablation Experiment

  • Removing the LSTM slows convergence and lowers final rewards; PPO-only training eventually reaches comparable reward but converges substantially more slowly.Replacing the pretrained LLM with a standard Transformer also slows convergence and lowers final rewards; heuristic supervised initialization accelerates and stabilizes PPO improvement.Source: E. Ablation Experiment · pp. 11–12

D. Generalization to Unseen-City Scenarios

  • In the unseen San Francisco scenario at SNR = 2 dB, AGENS rewards are close to exhaustive search, and its high-throughput strategy reduces BER to nearly zero with slightly lower R.Exhaustive search achieves the highest R and maximum reward; AGENS uses a larger power compensation level.Source: D. Generalization to Unseen-City… · p. 11, Table IV · p. 12

F. Model Complexity

  • AGENS plans a complete strategy in one forward pass, reducing inference time relative to search-based heuristics, but is designed for slower link-configuration planning rather than slot-level reconfiguration.The strategy is generated for a channel condition and user intent, then maintained over an operation interval.Source: F. Model Complexity · p. 12, Table V · p. 12

G. Interactive Intent-Aware Strategy Generation

  • In the interaction example, AGENS interprets multimodal environmental and transmission inputs and generates an interpretable PHY configuration aligned with a high-throughput preference.The example relates the preference to services such as video streaming or cloud gaming.Source: G. Interactive Intent-Aware Strategy… · p. 12, Figure 9 · p. 12

VII. CONCLUSION

  • AGENS selects from predefined PHY-module candidates to remain compatible with modular systems, and the paper reports competitive link-level performance with lower inference latency than search-based heuristics.The conclusion characterizes it as adaptable across heterogeneous channels, user preferences, and unseen-city scenarios.Source: §VII · pp. 12–13
Loading 2610.34131v1…