WirelessLLM-Agent: A Unified LLM-Based Agent Framework for Multi-Task Wireless Communication Decision-Making

The integration of large language models into wireless communication has shown promising results for individual tasks. However, existing approaches are typically designed for single-task scenarios and rely on supervised fine-tuning that fails to optimize for long-term decision quality. In this paper, we propose WirelessLLM-Agent, a unified LLM-based agent framework for multi-task wireless communication decision-making. Our framework integrates a semantic state serialization module that transforms heterogeneous wireless states into structured textual representations, a multi-task adapter architecture based on MoE-LoRA for parameter-efficient knowledge sharing, and a two-stage training paradigm combining SFT warm-start with GRPO reinforcement learning enhanced by lookahead collaborative simulation. Extensive experiments on channel multi-task learning, mobile edge computing task offloading, and cooperative edge caching demonstrate that WirelessLLM-Agent consistently outperforms existing methods while exhibiting strong zero-shot generalization.