Wireless Power Control Based on Large Language Models

TL;DR

This paper investigates the power control problem in wireless networks by repurposing pre-trained large language models (LLMs) as relational reasoning backbones by proposing PC-LLM, a physics-informed framework that augments a pre-trained Transformer with an interference-aware attention bias.

Abstract

This paper investigates the power control problem in wireless networks by repurposing pre-trained large language models (LLMs) as relational reasoning backbones. In hyper-connected interference environments, traditional optimization methods face high computational cost, while standard message passing neural networks suffer from aggregation bottlenecks that can obscure critical high-interference structures. In response, we propose PC-LLM, a physics-informed framework that augments a pre-trained Transformer with an interference-aware attention bias. The proposed bias tuning mechanism injects the physical channel gain matrix directly into the self-attention logits, enabling explicit fusion of wireless topology with pre-trained relational priors without retraining the backbone from scratch. Extensive experiments demonstrate that PC-LLM consistently outperforms both traditional optimization methods and state-of-the-art graph neural network baselines, while exhibiting exceptional zero-shot generalization to unseen environments. We further observe a structural-semantic decoupling phenomenon: Topology-relevant relational reasoning is concentrated in shallow layers, whereas deeper layers encode task-irrelevant semantic noise. Motivated by this finding, we develop a lightweight adaptation strategy that reduces model depth by 50\%, significantly lowering inference cost while preserving state-of-the-art spectral efficiency.

Figures (7)

• Figure 1: Graph representation of the wireless interference network. Communication links within transceiver pairs are modeled as nodes, while the mutual interference between pairs is represented by directed edges, constituting a fully connected interference graph $\mathcal{G}$.
• Figure 2: Overall architecture of the proposed PC-LLM. The framework aligns physical channel features with the latent space of the model through an input projection layer, a bias-modulated Transformer backbone, and a power inference head. Red modules denote trainable components including task-specific layers initialized from scratch and pre-trained layers adapted via LoRA, while blue modules represent the frozen backbone.
• Figure 3: Performance comparison of the proposed PC-LLM framework against baseline algorithms across varying network densities (number of D2D pairs $K$) and interference environments (distance ranges $[d_{\min}, d_{\max}]$). The results are organized by optimization objectives: (a)–(c) sum-rate maximization, (d)–(f) proportional fairness, and (g)–(i) harmonic maximization. The performance is normalized with respect to the WMMSE-Best algorithm for the max sum rate and proportional fairness tasks, whereas the harmonic mean rate is normalized relative to PCGNN-Large-Multi.
• Figure 4: Performance comparison of the proposed PC-LLM and baseline algorithms under multi-dataset and single-dataset training configurations. The performance is normalized with respect to the WMMSE-Best algorithm for the max sum rate and proportional fairness tasks, whereas the harmonic mean rate is normalized relative to PCGNN-Large-Multi. The bars represent the average performance across 15 distinct scenarios with varying user densities and channel distributions, while the error bars denote the standard deviation, indicating the robustness of each method.
• Figure 5: Normalized generalization performance of multi-dataset trained models tested on the unseen wide-range scenario ($[d_{\min}, d_{\max}] = [1, 100]$ m) with varying user densities $K$. The performance is normalized with respect to the WMMSE-Best algorithm for the max sum rate and proportional fairness tasks, whereas the harmonic mean rate is normalized relative to PCGNN-Large-Multi.