OAT uses one-class learning to trace agentic failure from success flows

Researchers propose OAT, a lightweight model for unsupervised failure attribution in LLM-based agentic systems that trains exclusively on successful trajectories. By casting the problem as one-class learning with neural controlled differential equations, OAT identifies error steps at inference time based on their deviation from learned dynamics.

• Trains on only 100 successful trajectories without step-level supervision.
• Assigns anomaly scores to failure trajectory steps based on latent space deviations.
• Achieves +20% and +7% F1 score improvements over prompting-based baselines in-domain and out-of-distribution, respectively.
• Runs 200–5000× faster than existing prompting-based approaches.

OAT provides a promising and efficient direction for diagnosing agentic system failures by eliminating the need for costly error annotations.