RADAR: Redundancy-Aware Diffusion for Multi-Agent Communication Structure Generation

Researchers have introduced RADAR, a novel generative framework designed to optimize communication structures in large language model-based multi-agent systems. While multi-agent systems excel in tasks like code generation and reasoning, their performance is often hindered by fixed or single-step communication topologies, leading to inefficient token usage and limited adaptability. RADAR addresses this by employing conditional discrete graph diffusion models to create redundancy-aware and query-adaptive communication topologies through a step-by-step generation process. This approach allows for fine-grained structural exploration, actively reducing communication overhead. Comprehensive experiments across six benchmarks demonstrate that RADAR consistently outperforms existing baselines, achieving higher accuracy, lower token consumption, and enhanced robustness in diverse scenarios. The study highlights the importance of dynamic topology design in improving the efficiency and effectiveness of multi-agent interactions. The associated code and data have been made publicly available to support further research and development in this area.