MARS repairs failed multi-agent runs via Monte-Carlo Tree Search, plus StateMAS: 1,310 replayable failure trajectories

arXiv·medium signalXBlueskyLinkedInCopy link

Hanxiao Lu and Tianyi Zhang posted MARS on July 31 (arXiv 2607.29055), targeting the gap after failure attribution: once you know which agent went wrong in a multi-agent trajectory, nothing automates the fix. MARS formulates repair as an MCTS process with diagnosis-guided expansion and taxonomy-augmented evaluation, and — critically for cost — scores candidate repairs with partial rollout instead of full simulation. They also release StateMAS, a benchmark of 1,310 replayable multi-agent failure trajectories across four agent architectures and four LLM backbones. MARS beats prior methods by 3.0 to 12.1 absolute points at comparable token cost; the replayable-failure corpus is arguably the more reusable artifact for anyone building agent evals.