aiqbench analysis
This paper introduces four rigorous tests to distinguish whether LLM judges genuinely verify faulty steps or merely rely on spurious surface cues in failure-attribution benchmarks. By demonstrating that simple predictors often match or exceed current model performance on these tasks, the authors highlight critical vulnerabilities in how we evaluate reasoning robustness and attribution accuracy. The work provides essential methodological guidance for developing more reliable benchmarks that prevent models from exploiting superficial patterns rather than performing actual cross-step verification.
Auditing Failure Attribution: Label Predictability and Judge Behavior
arXiv:2610.17747v1 Announce Type: cross Abstract: Failure-attribution benchmarks ask an LLM judge which step or agent caused a failure. A high score should mean the judge checked the faulty step against the run (a cross-step check). However, spurious signals (surface cues) that predict the annotation can produce a high score. In this paper, we propose four tests to separate checking from cue use. Input restriction asks whether partial input predicts the annotation. Cue following checks whether moved cues change answers. Residual accuracy scores judges where cue-only predictors fail. Aggregation asks whether the failure mix or metric changes rankings. Real observations motivate these tests: on three benchmarks, simple predictors match or beat GPT-4.1. In two Who&When Pro systems, a classifier reading one step's reasoning finds 95-100% of faulty steps, motivating input restriction. In one of them, error injection leaves no reasoning after the faulty step; judges fine-tuned on its labels beat GPT-4.1 by 41 points. Adding task-related reasoning after that step moves this cue: fine-tuned judges lose 31-99 points, but zero-shot judges lose at most 9, motivating cue following. On Aegis, blaming every agent matches GPT-4.1 on micro-F1 but trails by 19 points in exact-set accuracy, motivating aggregation. We recommend reporting these tests when developing failure-attribution benchmarks.
Abstract:Failure-attribution benchmarks ask an LLM judge which step or agent caused a failure. A high score should mean the judge checked the faulty step against the run (a cross-step check). However, spurious signals (surface cues) that predict the annotation can produce a high score. In this paper, we propose four tests to separate checking from cue use. Input restriction asks whether partial input predicts the annotation. Cue following checks whether moved cues change answers. Residual accuracy scores judges where cue-only predictors fail. Aggregation asks whether the failure mix or metric changes rankings. Real observations motivate these tests: on three benchmarks, simple predictors match or beat GPT-4.1. In two Who&When Pro systems, a classifier reading one step's reasoning finds 95-100% of faulty steps, motivating input restriction. In one of them, error injection leaves no reasoning after the faulty step; judges fine-tuned on its labels beat GPT-4.1 by 41 points. Adding task-related reasoning after that step moves this cue: fine-tuned judges lose 31-99 points, but zero-shot judges lose at most 9, motivating cue following. On Aegis, blaming every agent matches GPT-4.1 on micro-F1 but trails by 19 points in exact-set accuracy, motivating aggregation. We recommend reporting these tests when developing failure-attribution benchmarks.
Computation and Language (cs.CL); Artificial Intelligence (cs.AI); Machine Learning (cs.LG)
arXiv:2610.17747 [cs.CL]
(or arXiv:2610.17747v1 [cs.CL] for this version)
https://doi.org/10.48550/arXiv.2610.17747 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 301 words
Related coverage
In open RLVR, “improvement” depends on the instrument — a small GRPO testbed separating what training optimizes, measures, and teaches
LessWrong · Jun 15, 2026
[AINews] not much happened today
Latent Space · Sep 18, 2026
Models don’t seem to be dishonest in the way humans are
LessWrong · Jul 22, 2026
The J-Space Debate, Agent Swarms, and Pacing Frontier AI - Digital Minds Newsletter #4
LessWrong · Sep 18, 2026