MARS Framework Automates Multi-Agent System Error Repair

Multi-agent systems (MAS) are becoming increasingly common for complex tasks, but identifying and fixing errors in their outputs remains a manual, labor-intensive process. Current methods focus on attributing failures, but automated recovery mechanisms are largely undeveloped. This new research proposes MARS, a novel framework that frames MAS repair as a Monte Carlo Tree Search (MCTS) problem.MARS navigates the vast space of potential repairs using diagnosis-guided expansion and taxonomy-augmented evaluation. Unlike traditional MCTS, it employs partial rollouts to reduce token consumption, making it more efficient. The team also created StateMAS, a comprehensive benchmark with 1,310 replayable multi-agent failure trajectories across various agent architectures and LLM backbones.Experiments on StateMAS show that MARS consistently surpasses state-of-the-art methods, achieving an absolute performance improvement of 3.0% to 12.1% while maintaining comparable token costs. Key components like taxonomy-augmented evaluation and diagnosis-guided expansion were confirmed as crucial for these gains.