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

За 30 секунд

Что произошло

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

Почему это важно

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.

На что обратить внимание

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