Tracing Agentic Failure from the Flow of Success

arXiv cs.CL · 18 hours ago

Researchers introduced OAT, a method for identifying which steps cause failures in LLM-based agent systems by training exclusively on successful trajectories using neural controlled differential equations. The approach achieved 200-5000 times faster performance than prompting-based baselines while improving F1 scores by 20% on in-domain and 7% on out-of-distribution datasets. This enables efficient debugging of agentic systems without requiring costly step-level error annotations on failure data.