Detect Before You Attribute: Cascade Failure Attribution for Multi-Agent Systems
DUOTRACE narrows long agent trajectories before asking an LLM to assign blame.
The paper proposes a plug-and-play filter for failure attribution in multi-agent systems. It detects anomalous executions first, then passes focused trajectory evidence to existing LLM-based attribution methods. The authors say this helps with long-context degradation while preserving semantic and structural signals from agent runs. Across six LLM-based baselines, DUOTRACE improved agent-level attribution accuracy by 8.7% and step-level accuracy by 7.0%. ArXiv · AI/CL/LG's note
The paper proposes a plug-and-play filter for failure attribution in multi-agent systems. It detects anomalous executions first, then passes focused trajectory evidence to existing LLM-based attribution methods. The authors say this helps with long-context degradation while preserving semantic and structural signals from agent runs. Across six LLM-based baselines, DUOTRACE improved agent-level attribution accuracy by 8.7% and step-level accuracy by 7.0%. ArXiv · AI/CL/LG's note
score 4