TRAJDEBUG: Tracing Error Lifecycle to Identify Critical Failures in Long-Horizon Agent Trajectories

In 30 Sekunden

Was passiert ist

TrajDebug framework identifies critical errors in long multi-step LLM agent trajectories by tracing error lifecycle and determining which failures caused final task failure.

Warum es zählt

Engineers building or debugging LLM-based agents need this when multi-step tasks fail and they must pinpoint which early mistake cascaded into the final error.

Achtung

The benchmark contains only 486 manually annotated trajectories from two sources, so performance on other agent types or domains remains unclear.