AgenticPrecoding: LLM-Empowered Multi-Agent System for Precoding Optimization

Precoding is essential for interference management in multi-antenna wireless systems, but traditional methods are tailored to specific models, limiting adaptability to future 6G networks. To address this, researchers from Zhejiang University developed AgenticPrecoding, a universal multi-agent framework that automates end-to-end precoding directly from user-level communication requirements. The system decomposes the derivation into four coordinated stages: problem formulation, solver selection, prompt upsampling, and code generation. Each stage is assigned to a specialized agent: two LoRA-adapted reasoning agents inject precoding-specific domain knowledge for formulation and solver selection, while two general-purpose LLMs handle prompt refinement and executable code generation.

A feedback-driven refinement mechanism enhances code executability, constraint feasibility, and solution quality. In extensive experiments across 10 representative precoding scenarios, AgenticPrecoding demonstrated superior cross-scenario adaptability compared to conventional optimization-based and LLM-based baselines. The framework essentially turns abstract communication goals into working MATLAB code, automating what currently requires expert manual tuning. By combining domain-specific fine-tuning (LoRA) with general-purpose LLM capabilities, AgenticPrecoding points toward a future where AI autonomously designs key components of wireless infrastructure.