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

A new framework called RADAR (Redundancy-Aware and query-adaptive generative framework) has been introduced to optimize communication topologies in multi-agent systems based on large language models. Unlike traditional fixed or single-step topologies, RADAR uses conditional discrete graph diffusion models to generate communication structures step-by-step, guided by the effective size of the graph. This approach reduces communication overhead by avoiding excessive token usage on simple tasks while enhancing capability on complex tasks. The method addresses limitations in current multi-agent systems, which often rely on rigid communication structures that hinder fine-grained exploration and flexible composition. RADAR is motivated by recent progress in graph diffusion models and aims to improve both efficiency and robustness across diverse tasks such as code generation, mathematical reasoning, and planning.