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

Abstract:

Large language model-based multi-agent systems have demonstrated remarkable proficiency across a wide array of applications, such as planning, mathematical reasoning, and code generation, often surpassing the capabilities of individual agents. However, the robustness and efficacy of these systems are fundamentally dependent on their communication topology. Currently, these topologies are typically static or constructed in a single step, which hinders fine-grained structural exploration and flexible composition. This limitation leads to inefficient token usage on simpler tasks and constrains performance on more complex ones.

To address this issue, we propose RADAR, a generative framework that is both redundancy-aware and query-adaptive, designed to actively minimize communication overhead. Drawing inspiration from recent advancements in conditional discrete graph diffusion models, we frame the design of communication topologies as an iterative, step-by-step generation process, with the effective size of the graph serving as the guiding metric. Our comprehensive evaluations across six benchmarks reveal that RADAR consistently surpasses recent baseline methods. It delivers superior accuracy, reduced token consumption, and enhanced robustness across various scenarios. The source code and data are publicly accessible at https://github.com/cszhangzhen/RADAR.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC