TRAJDEBUG: Tracing Error Lifecycle to Identify Critical Failures in Long-Horizon Agent Trajectories
TrajDebug tries to find the first failure that actually caused a long agent run to collapse.
The paper frames the problem as critical error detection in failed LLM-agent trajectories, where many local mistakes may appear but only some drive the final outcome. Its method compresses long histories at multiple levels, identifies errors from evidence, then tracks whether each error was resolved and whether it still affected the terminal failure. The authors also introduce TrajErrBench, with 486 manually annotated failed trajectories from Tau2Bench and SWE-Bench Pro. They report best overall performance against baselines and say the diagnoses can give actionable feedback for improving downstream agent success. ArXiv · AI/CL/LG's note
The paper frames the problem as critical error detection in failed LLM-agent trajectories, where many local mistakes may appear but only some drive the final outcome. Its method compresses long histories at multiple levels, identifies errors from evidence, then tracks whether each error was resolved and whether it still affected the terminal failure. The authors also introduce TrajErrBench, with 486 manually annotated failed trajectories from Tau2Bench and SWE-Bench Pro. They report best overall performance against baselines and say the diagnoses can give actionable feedback for improving downstream agent success. ArXiv · AI/CL/LG's note
score 5