LLM-enabled Antenna Partitioning and Beamforming Optimization for Segmented Pinching

Abstract: Integrated sensing and communication (ISAC) requires spatial architectures that can flexibly balance data transmission and environment sensing. Segmented pinching antenna-assisted ISAC provides such flexibility by allowing different waveguide segments to be dynamically configured for transmission and reception. However, its design involves the joint optimization of antenna deployment, segment partitioning, and beamforming under coupled communication and sensing constraints, which becomes particularly challenging when the numbers of communication users and sensing targets vary across scenarios. To endow the system with stronger adaptability to changing user and target configurations, 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 produce permutation-invariant representations of user-target interactions, and the resulting features are processed by a LLM backbone with two task-specific heads for jointly predicting antenna deployment, segment partitioning, and ISAC beamforming. In addition, a user count transfer mechanism is developed to examine whether the learned deployment policy is site-specific and reusable under changed user configurations. Simulation results show that the proposed framework achieves higher communication rates while maintaining reliable sensing accuracy. Moreover, the learned deployment policy remains highly stable when transferring to other user counts, which reduces the training cost from full model retraining to beamforming head adaption.

• The paper presents a learning-based framework combining SGNN and an LLM with LoRA adaptation to jointly optimize antenna deployment, partitioning, and beamforming.
• It outperforms traditional MLP and transformer baselines by achieving faster convergence, higher sum rates, and robust sensing under dynamic configurations.
• The study demonstrates efficient transferability by decoupling site-specific deployment from user-specific beamforming, significantly reducing retraining overhead.

LLM-enabled Antenna Partitioning and Beamforming Optimization for Segmented Pinching Antenna-Aided ISAC

Introduction and Motivation

The integrated sensing and communication (ISAC) paradigm has expanded the demands on antenna architectures, necessitating spatial resource allocation that simultaneously supports downlink data throughput and environmental awareness. Segmented pinching antenna (SWAN) systems offer distinctive architectural flexibility by enabling independent control of waveguide segments and reconfigurable assignment of transmit and receive modes to specific aperture regions. However, joint optimization of antenna deployment, segment-wise partitioning, and ISAC-compatible beamforming is inherently a high-dimensional mixed discrete-continuous problem, further complicated by dynamic user and target configurations.

The paper addresses these challenges by proposing a learning-based framework that leverages permutation-invariant, graph-structured channel representations and a LLM backbone, aiming to deliver robust ISAC performance and transferability across heterogeneous scenarios.

System Model and Problem Formulation

The SWAN-ISAC system comprises a base station (BS) with $M$ controllable dielectric waveguide segments, $N$ reconfigurable pinching antennas, $K_c$ communication users, and $K_s$ sensing targets. The architecture allows non-uniform antenna deployment and segment-wise reallocation between transmit and receive modes.

Key aspects include:

• Antenna Deployment Optimization: Physical placement of PAs along the waveguide, subject to minimum inter-element distances and aperture constraints.
• Segment-wise Tx/Rx Partitioning: Assigning each segment to either transmit or receive mode, with the partitioned segments forming effective communication and sensing apertures.
• Beamforming in Near-Field Regime: Jointly optimized communication and probing beams must adhere to SWAN's structural constraints and satisfy SINR and sensing accuracy objectives.
• Joint Objective: Maximize aggregate communication rates under stringent CRLB-based localization error constraints for all sensing targets.

This results in a nontrivial constrained optimization task, with strong coupling between physical placement, segment assignment, and high-dimensional beamforming vectors.

Learning Framework Architecture

To overcome scalability barriers in traditional model-based or supervised deep learning approaches—especially under variable user/target counts—the authors introduce a composite learning architecture with four principal components:

1. CSI-Induced Self-Graph Encoder: Constructs a variable-sized, permutation-invariant graph representation, encoding inter-user/target relationships via adaptive adjacency matrices derived from channel similarity metrics.
2. LLM Backbone with LoRA Adaptation: A pretrained GPT-2 style LLM is repurposed for sequence modeling of antenna-domain channel-state tokens. LoRA (low-rank adaptation) enables lightweight, parameter-efficient fine-tuning while freezing the majority of backbone weights.
3. Task-Specific Output Heads: Decoupled heads handle (a) hierarchical deployment and segment-assignment predictions and (b) per-user and per-target beamforming.
4. User-Count Transfer Mechanism: Exploits the decoupling of site-specific deployment and user-specific beamforming. Under variations in user count, the deployment/partitioning pipeline is reused invariantly, and only the beamforming head is retrained, drastically reducing adaptation effort.

The joint loss comprises deployment supervision, communication rate maximization, sensing penalty (via log-scale hinge losses on CRLB violations), and geometric regularization for feasible apertures.

Numerical Results and Empirical Claims

The authors report several strong empirical findings substantiated by extensive simulation:

• Superior Performance: The proposed SGNN_LLM_SH model (incorporating self-graph encoding, LLM backbone, and split heads) consistently outperforms MLP, transformer, and plain LLM baselines in terms of validation loss, sum rate, and sensing CRLB across all communication–sensing power splits.
• Convergence Stability: SGNN_LLM_SH exhibits the fastest and most robust convergence profile, avoiding the instability seen in non-graph-based models.
• Deployment Robustness: Learned deployment strategies are stable and physically meaningful, consistently producing segment-wise allocations compatible with hybrid transmission/sensing requirements.
• Robustness to Power Split and CSI Mismatch: Model performance degrades gracefully under reduced communication power and increasing CSI perturbation, maintaining superiority to all baselines.
• Transferability: When adapting to new user/target configurations, the decomposition between deployment/partition and beamforming is validated empirically: by freezing the deployment head and retraining only the beamforming head, the framework attains near-source task performance with two orders of magnitude fewer trainable parameters and substantially reduced wall-clock adaptation time.
• Model Efficiency: While LLM-based models are computationally heavier than lightweight baselines, the incremental complexity due to self-graph encoding is modest relative to LLM parameter scale, and is offset by pronounced gains in performance and adaptation nimbleness.

Theoretical and Practical Implications

• Hierarchical Decomposition: The separation of site-dependent structural design from user-dependent beamforming introduces an avenue for site-specific policy reuse—critical for future 6G and ISAC deployments where operational scenarios shift rapidly.
• Learning-based ISAC: The success of LLMs, when adapted to structured physical-layer inputs via permutation-invariant pre-processing, suggests broader applicability of foundation models in complex physical systems, provided they are coupled to domain-aware data encoders.
• Parameter-Efficient Transfer: The empirical demonstration of low-rank adaptation (LoRA) in non-NLP wireless domains paves the way for edge-side resource-aware model adaptation under changing network conditions.

Future Prospects

The framework has considerable potential extensions: online adaptation to user/target mobility, domain adaptation under hardware impairments or sparse training data, incorporation of hardware-aware constraints (e.g., quantized phase control), and real-time execution in SWAN-ISAC testbeds. Furthermore, the general recipe of graph-enhanced LLMs with LoRA is transferable to other wireless optimization scenarios, including reconfigurable intelligent surfaces and ultra-massive ISAC systems.

Conclusion

The study meticulously formulates and solves the coupled deployment, partitioning, and beamforming optimization problem in segmented pinching antenna-based ISAC, overcoming the critical challenge of variable scenario generalization. Through a hybrid SGNN–LLM framework with LoRA adaptation, the method achieves high communication rates, robust sensing, and scalable low-overhead transfer across user configurations. The results substantiate the viability of LLM-driven design in next-generation ISAC, with both theoretical insight into architecture decomposition and practical efficacy for rapid, hardware-realizable optimization (2604.10372).