Unsupervised failure detection for LLM agents trained on success data
Debugging agentic AI systems at scale has hit a wall: existing failure attribution methods either burn compute through prompting or demand expensive hand-labeled error trajectories. A new paper proposes OAT, a lightweight one-class learning approach that trains exclusively on successful agent runs, then identifies failure points at inference time without step-level supervision. This shifts the economics of agent reliability from annotation-heavy to data-efficient, potentially unlocking faster iteration cycles for teams building production LLM agents where failure diagnosis currently remains a bottleneck.
Modelwire context
ExplainerThe key insight is inverting the debugging problem: instead of labeling failure trajectories (expensive) or querying models about failures (compute-heavy), OAT learns what success looks like, then flags deviations at inference time. This is a methodological flip, not just an efficiency gain.
This connects directly to the broader agent reliability infrastructure emerging across recent work. The co-evolving metrics paper from July tackled how to evaluate self-improving agents without human-in-the-loop validation; OAT solves the complementary problem of diagnosing why agents fail once deployed. Similarly, MemOps exposed hidden failure modes in agent memory management, while Deep4ge built fault detection into training trajectories. OAT extends this trajectory-analysis pattern to inference, completing a diagnostic chain from training through deployment. The common thread: moving from binary pass/fail scoring to systematic failure attribution.
If OAT's one-class approach maintains accuracy parity with supervised baselines on held-out agent tasks from real production systems (not just benchmark environments), that validates the core claim. Watch whether the authors release code and whether teams at Anthropic, OpenAI, or other agent-builders adopt it within six months; adoption velocity will signal whether the compute savings actually matter at scale.
Coverage we drew on
This analysis is generated by Modelwire’s editorial layer from our archive and the summary above. It is not a substitute for the original reporting. How we write it.
MentionsOAT
Modelwire Editorial
This synthesis and analysis was prepared by the Modelwire editorial team. We use advanced language models to read, ground, and connect the day’s most significant AI developments, providing original strategic context that helps practitioners and leaders stay ahead of the frontier.
Modelwire summarizes, we don’t republish. arXiv cs.CL originally reported this story as “Tracing Agentic Failure from the Flow of Success”. The full content lives on arxiv.org. If you’re a publisher and want a different summarization policy for your work, see our takedown page.