OAT: Finding Failed Steps via Successful Trajectories

Samuel Yeh · hf · 2026-07-16

This work investigates **failure attribution for LLM agents**: how to identify the exact step that causes task failure within a failed trajectory. - **Existing Issues**: - Prompt-based methods are computationally expensive; - Methods relying on step-by-step annotation of failed trajectories are costly and hard to scale. - **Core Approach**: Proposes **OAT**, an unsupervised method trained exclusively on successful trajectories to locate failed steps during inference. - Formulates the problem as **one-class learning**; - Uses **neural controlled differential equations** to model the dynamic patterns of successful trajectories in latent space; - During inference, assigns anomaly scores to steps in failed trajectories based on their deviation from "success dynamics," forming an error step set. - **Results**: Trained on just **100 successful trajectories**, OAT is **200–5000x faster** than prompting-based baselines while achieving **+20%** and **+7%** F1 improvements on in-domain and out-of-distribution datasets, respectively. - **Significance**: Provides a lightweight, scalable failure diagnosis direction for agent debugging.

Related event: Microsoft Proposes OAT for LLM Agent Failure Attribution(2 posts)→

Original post →

More from coding & agent

coding & agent channel →