Locating Hidden Failures Makes Long-Horizon Agents More Reliable
Salman Rahman, Yubin Kim, Mihir Parmar, A. Ali Heydari, Genglin Liu, Simon A. Lee, Weizhi Zhang, Arian Hosseini, et al.
cs.LG
Quick verdict
Human-verified benchmark shows frontier judges cannot locate the first error in long agent runs while a 4B trained verifier can - largely well supported, with draft-state artifacts and LLM-judged taxonomy as caveats.
Summary
The paper studies how long-horizon LLM agents fail by annotating 2,518 real trajectories from SWE-bench, TerminalBench, and BixBench, spanning six model families and multiple harnesses. Two human annotators label every step and mark each failed run's first mistake (Cohen's kappa 0.77 on failure, 0.73 on location); 6,967 mistakes are typed against a pre-registered 88-category taxonomy by LLM judges. Reported signature: agents recover from the first mistake in only 30.5% of runs and never self-detect it in 38.5%; runs typically end looking correct; reasoning/planning faults originate failures while action errors are downstream cascades (next-step error probability jumps 3-5% to 40-58%). They release Traverse, a 1,423-trajectory/44,341-step benchmark for first-mistake localization and outcome verification, on which six frontier judges exact-match the first mistake on under a third of long runs (best 26.8% SWE-bench, 32.3% TerminalBench) and all show detection bias. Scout, a 4B Qwen3 model trained with SFT then GRPO using a recall-product reward, reportedly beats these judges, transfers to unseen BixBench, and raises best-of-N success (81.8% to 90.2% on Terminal-Bench 2.0; 75.6% to 78.0% on SWE-bench Verified). Main limitations: fine-grained taxonomy typed by LLM judges with kappa 0.30, draft-state artifacts and numeric inconsistencies in the text, and best-of-N comparisons against only two verifiers without uncertainty estimates.
Review complete
Assessment at a glance
6
risk flags identified
● high 0
● medium 4
● low 2
5 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Long-horizon agent failure follows a recurring signature: after the first mistake agents fail to recover in 69.5% of runs, never self-detect in 38.5%, and 84% of failed runs end on a locally correct-looking step.
- Reasoning and planning faults are the originating causes of failure while action errors are mostly downstream cascades; one wrong step makes the next roughly 11-14x likelier to be wrong.
- Task success does not imply safety: runs scored as solved include irreversible destructive actions and fabricated success, most of which escape regex-level checks.
- No frontier judge locates the first mistake on more than a third of long-horizon runs (best: 26.8% SWE-bench, 32.3% TerminalBench exact match), and every judge is systematically biased at trajectory-level error detection.
- Scout, a 4B verifier trained with SFT plus GRPO on the annotations, beats these judges at localization, transfers to an unseen science domain, and raises best-of-N success above the agents' own pass@1 (e.g., 81.8% to 90.2% on Terminal-Bench 2.0) without retraining the agent.
Strengths
Large, human-verified step-level ground truth
1,423 trajectories / 44,341 steps in Traverse with dual independent annotation (kappa 0.77 failure, 0.73 location) and third-annotator adjudication; a scale of human step-labeling rare in agent evaluation.
Well-positioned benchmark contribution
Table 5 shows Traverse is the only process-verifier benchmark combining 20-200-step horizons, naturally occurring (not synthetic) failures, human first-mistake localization, multiple scaffolds, and outcome-level grading in one resource.
Unusually candid methodology
Principled reward design for the verifier
Novel safety audit of solved runs
Weaknesses
The headline taxonomy is LLM-judged, not human-judged
Draft-state artifacts and numeric inconsistencies
Best-of-N evaluation is narrow and unquantified
First-mistake framing excludes a third of the failure space
Judge protocol may understate frontier capability
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 6 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is the Traverse benchmark?
Traverse is a released benchmark of 1,423 real LLM-agent trajectories (44,341 steps) from software engineering, computer-use, and science tasks. Every step is human-labeled correct or incorrect and the first mistake of each failed run is marked, so models can be graded on whether they can find where a long agent run first went wrong and whether a run failed at all.
What is Scout and how does it work?
Scout is a 4-billion-parameter verifier (built on Qwen3-4B-Thinking) trained to read an agent trajectory step by step and flag the first error. It is first fine-tuned on trajectories with human labels plus label-faithful chain-of-thought analyses, then refined with reinforcement learning (GRPO) using a reward that only pays off when the model both leaves correct steps alone and catches real mistakes. The authors report it beats much larger frontier models at localization and improves best-of-N agent selection.
Why do long-horizon AI agents fail?
According to this study, they usually do not fail from many independent mistakes but from one early mistake that goes unnoticed and uncorrected: agents recover in only about 30% of runs after their first error, and about 84% of failed runs end looking correct. The originating errors are typically reasoning and planning faults; the many invalid or repeated actions seen later are mostly downstream fallout.
Can frontier LLMs detect when an AI agent made a mistake?
Partially. On this paper's benchmark, the best frontier judges find the exact first-mistake step in fewer than a third of long trajectories (26.8% on software engineering, 32.3% on terminal tasks), though they are usually within a few steps. Telling whether a run failed at all is easier (F1 up to about 88%), but every judge is biased, either over-flagging correct runs or missing real failures.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- The extracted text is truncated: Scout's quantitative localization/detection numbers on the held-out set are not present (the Table 2 Scout row is garbled), so the central Scout-beats-judges margin could not be verified from this text.
- The text contains draft-state artifacts (Appendix ??, [Extended Data Fig. X], [pending the recovery and cross-domain analyses]) and at least two internal numeric inconsistencies (2,620 vs 2,518 trajectories; silent-failure 84.1% vs 83.5%), so the reviewed version may not match the final release.
- Figures were assessed from captions and embedded labels only; per-figure data quality could not be independently checked.
- Appendix content (annotation guidelines, per-category definitions beyond Table 3, extended data) was not available in the extraction.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.