MARS: Automated Repair for Multi-Agent Systems via Monte-Carlo Tree Search
A novel framework called MARS (Multi-Agent Repair via Search) has been introduced to facilitate the automation of repairing multi-agent systems (MAS) that yield incorrect or inadequate results. Typically, users need to manually pinpoint errors by analyzing agent trajectories (failure attribution) and then give feedback for output refinement (repair). Although recent advancements have tackled failure attribution, automated repair has not been significantly explored. MARS approaches MAS repair through a Monte Carlo Tree Search (MCTS) method, exploring the extensive repair possibilities via diagnosis-driven expansion and taxonomy-enhanced evaluation. In contrast to conventional MCTS, which relies on full rollouts for evaluation, MARS employs partial rollouts to minimize token usage. This framework is presented alongside StateMAS, a comprehensive benchmark featuring 1,310 replayable multi-agent scenarios. The research paper can be found on arXiv with the identifier 2607.29055, categorized as 'cross'. This study fills a vital gap in MAS deployment, paving the way for automated error recovery in intricate multi-agent systems.
Key facts
- MARS is a search-based framework for automated repair of multi-agent systems.
- It formulates MAS repair as a Monte Carlo Tree Search (MCTS) process.
- MARS uses diagnosis-guided expansion and taxonomy-augmented evaluation.
- It employs partial rollouts to reduce token consumption compared to standard MCTS.
- StateMAS is a new benchmark with 1,310 replayable multi-agent scenarios.
- The paper is available on arXiv with identifier 2607.29055.
- The announcement type is 'cross'.
- Automated repair mechanisms for MAS have been largely unexplored.
Entities
Institutions
- arXiv