MARS framework auto-repairs multi-agent failures with 12.1% accuracy jump
Tree search finds and fixes AI agent mistakes across 1,310 failure trajectories.
Multi-agent systems (MAS) are powerful but brittle: when they fail, engineers must manually trace agent trajectories to pinpoint mistakes and craft repair feedback. A new paper from Hanxiao Lu and Tianyi Zhang introduces MARS, a search-based framework that fully automates this repair loop. Instead of relying on human intervention, MARS treats repair as a Monte Carlo Tree Search problem, exploring the vast space of possible corrections and zeroing in on effective fixes.
What makes MARS stand out is its efficiency. Standard MCTS evaluates complete simulations via full rollouts, but MARS uses partial rollouts on agent trajectories, dramatically reducing token consumption. It also uses diagnosis-guided expansion to target likely failure points and taxonomy-augmented evaluation to score candidate repairs. To validate the approach, the authors built StateMAS, a large-scale benchmark with 1,310 replayable multi-agent failure trajectories across four agent architectures and four LLM backbones. Results show MARS consistently beats existing methods, achieving absolute improvements ranging from 3.0% to 12.1% across all settings while keeping costs on par with baselines—a significant step toward self-healing AI systems.
- MARS applies Monte Carlo Tree Search to automate failure attribution and repair in multi-agent systems.
- StateMAS benchmark includes 1,310 replayable failure trajectories across 4 agent architectures and 4 LLM backbones.
- MARS improves repair success by 3.0–12.1% over state-of-the-art while keeping token consumption comparable.
Why It Matters
Turns multi-agent debugging from a manual chore into automated self-repair, cutting operational overhead and accelerating reliable AI deployment.