Graph-Enhanced LLM for SWAN-ISAC

Abstract: Segmented pinching antenna assisted integrated sensing and communication (ISAC) systems enable flexible spatial resource utilization by allowing different waveguide segments to be dynamically configured for transmission and reception. However, the resulting design requires the joint optimization of antenna deployment, segment partitioning, and beamforming under coupled communication and sensing constraints. In this paper, we propose a general learning framework for segmented pinching antenna assisted ISAC systems. Specifically, a channel state information (CSI)-induced self-graph is constructed to capture the scenario-dependent interactions among communication users and sensing targets. Based on the learned graph representation, a LLM backbone with low-rank adaptation (LoRA) is employed, followed by two task-specific output heads for antenna deployment and beamforming prediction, respectively. Simulation results show that the proposed framework achieves a favorable tradeoff between communication rate and sensing accuracy

• The paper introduces a hybrid graph-enhanced LLM framework that jointly optimizes deployment, partitioning, and beamforming, addressing mixed discrete-continuous challenges in segmented ISAC systems.
• It leverages CSI-induced self-graphs and LoRA-adapted LLMs to capture dynamic channel conditions and encode context-dependent user and target interactions.
• Results show superior convergence, enhanced sum rate, and minimized CRLB, underscoring an effective communication-sensing tradeoff for future 6G wireless networks.

Graph-Enhanced LLM Framework for Segmented Pinching Antenna ISAC Systems

Introduction and Motivation

Segmented pinching antenna systems with integrated sensing and communication (ISAC) capabilities offer a new approach to efficiently share spatial resources and hardware for sixth-generation (6G) wireless networks. The SWAN-ISAC framework, which relies on segmented waveguides outfitted with pinching antennas, enables highly flexible, segment-wise assignment of transmission and reception, creating a fine-grained control over the spatial degrees of freedom. The resulting system design involves the simultaneous optimization of antenna deployment, segment partitioning, and beamforming in an environment characterized by dynamic user and target configurations, as well as coupled performance requirements for communication and sensing.

Model-based optimization methods are challenged by the mixed discrete-continuous nature of these decisions and the inherently variable channel conditions and network topologies. Recent advancements in deep learning, particularly graph neural networks (GNNs) and LLMs, have shown the ability to learn complex mappings and encode structured dependencies present in modern wireless systems. This paper introduces a hybrid graph-enhanced LLM framework which leverages channel-state self-graphs in conjunction with LLMs adapted via low-rank adaptation (LoRA) to jointly optimize the primary design variables in SWAN-ISAC systems (2604.10256).

System Model and Problem Formulation

The system comprises a base station with $M$ segmented waveguides and $N$ pinching antennas, dynamically configured to support $K_c$ communication users and $K_s$ sensing targets. The architectural flexibility allows arbitrary assignment of segments to transmit or receive modes, further influenced by the physical deployment of the antennas.

The optimization targets:

• Antenna deployment: Continuous, subject to physical and mutual exclusion constraints.
• Segment-wise Tx/Rx partitioning: Discrete binary variables per segment.
• Beamforming vectors: Complex-valued, conditioned by transmit segment assignment and power constraints.

Communication utility is quantified via aggregate sum rate, computed from user-specific SINR, while sensing performance is measured using the Cramér–Rao lower bound (CRLB) on target 3D position estimates, derived from the Fisher information matrix (FIM). The overall objective is a joint maximization of sum rate while ensuring that the CRLB for every target remains below a predefined threshold, subject to deployment constraints and power budgets.

Notably, the problem is highly structured—decision variables are not independent, and the optimality of any variable's setting is context-dependent, influenced by both user/target positions and segment assignments.

Learning-Based Solution: Graph-Enhanced LLM Architecture

To address these challenges, the proposed solution entails:

1. CSI-induced Self-Graph Encoder: Nodes represent users and sensing targets, with edge weights capturing normalized CSI similarities. Node features include CSI norms, aggregate phase, and type indicators, forming an instance-specific, permutation-invariant relational encoding. This makes the representation robust to orderings and suitable for environments with dynamic and heterogeneously ordered participants.
2. LoRA-Adapted LLM Backbone: The graph-derived embeddings are injected as soft conditioning to a pretrained GPT-style LLM, whose core layers have been LoRA-adapted for parameter-efficient finetuning. Input tokenization occurs over antenna-domain CSI, and shared representations are processed by the LLM.
3. Split Task-specific Output Heads: Separate heads decode the shared embedding for the deployment/partitioning variables and beamforming coefficients, respectively. This modularizes the output prediction and provides inductive bias to better match the structural task decomposition of the SWAN-ISAC system.
4. End-to-End Loss Formulation: The total loss incorporates direct supervision on deployment, penalties for CRLB violations, rewards for high sum rate, and geometric regularization to ensure physical feasibility.

The training pipeline is fully differentiable, with candidate beamformers and deployments evaluated in a synthetic, physics-consistent simulation loop during optimization.

Numerical Analysis and Comparative Results

Simulations were conducted using a scenario with $M=4$ segments, $N=40$ antennas, $K_c=2$ users, and $K_s=1$ target. The presented benchmarks compare the proposed model (SGNN_LLM_SH) against MLP, Transformer-based, plain LLM, and ablated versions omitting key architectural components.

Crucial findings include:

• SGNN_LLM_SH achieves the lowest validation loss and converges faster than all baselines.
• Communication performance: The proposed approach attains the highest achievable sum rate, exceeding MLP and Transformer-based baselines by a wide margin and outperforming plain LLM approaches, underlining the necessity of graph-based relational feature encoding.
• Sensing performance: The CRLB for position estimation is minimized more effectively without sacrificing communication throughput, highlighting that rate improvements are not achieved at the expense of sensing utility.
• Deployment prediction: Decoupling deployment and beamforming heads provides a more effective allocation of representational capacity, leading to an improved tradeoff in structural accuracy and communication-sensing joint objective.
• Structural generalization: Models trained without graph-induced embeddings and task-specific heads display impaired generalization, particularly in heterogeneous or dynamically varying scenarios.

Strong claims in the results: The framework robustly achieves a superior communication-sensing tradeoff relative to existing deep learning baselines, particularly in scenarios with variable numbers and arrangements of users and sensing targets.

Theoretical and Practical Implications

By integrating CSI-induced self-graphs with LoRA-finetuned LLMs, the architecture captures context-dependent interactions and enables effective learning across mixed discrete-continuous, highly coupled decision spaces. Permutation-invariance addresses the practical constraint of variable user/target presence, a challenge for traditional fixed-input models. The split head design reflects the orthogonality of variable types—structural (deployment/partition) and functional (beamforming)—optimizing overall system performance rather than surrogate objectives.

Practically, this methodology facilitates near real-time deployment adaptation and beam synthesis in advanced ISAC deployments, addressing a critical barrier to scalable ISAC realization in 6G and beyond. The low-rank adaptation allows reuse of large pretrained models without requiring impractically large domain-specific datasets.

Theoretically, this work suggests a path forward for hybrid network architectures in communications research, combining symbolic/graphical models with powerful sequence-processing LLMs—a paradigm extensible to other structured wireless optimization problems.

Future Directions

Potential advances include:

• Scaling the architecture to more complex topologies (multi-cell, multi-antenna users, bidirectional ISAC).
• Extending to hardware-in-the-loop or real-time over-the-air validation.
• Investigating transfer and meta-learning extensions for rapid adaptation to environment changes.
• Exploring further modularity, such as hierarchical graph construction or spatial-temporal (multi-slot) modeling capabilities.

Conclusion

This study introduces a graph-enhanced, LLM-based framework for the joint optimization of deployment, partitioning, and beamforming in segmented pinching antenna ISAC systems. The integration of CSI-induced self-graph relational modeling with LoRA-adapted LLMs and task-specialized output heads enables robust learning of structured policies, substantially improving the tradeoff between communication rate and sensing accuracy across dynamic scenarios. The presented framework clarifies the promise of relational, modular architectures for advanced wireless resource optimization (2604.10256).