TRACE-MAS: Diagnosing and repairing collaboration failures in multi-agent systems
TRACE-MAS: Diagnosing and Repairing Collaboration Failures in Multi-Agent Systems
AI summary
TRACE-MAS is a fine-grained framework for diagnosing and repairing collaboration failures in LLM multi-agent systems. It learns normal collaboration patterns from benign trajectories, detects deviations in agent state changes, message passing, and task dependency evolution, locates the responsible component through source-aware attribution, and restores collaboration with minimal intervention. On five MAS benchmarks, root-cause localization accuracy improves by 60.2%, and task recovery is consistently better than strong MAS defenses and guardrails.