Auditing the Audit: Five Failure Modes in Benchmark-Validity Audits

• URL: http://arxiv.org/abs/2607.02586v1
• Date: Wed, 01 Jul 2026 05:30:07 GMT
• Title: Auditing the Audit: Five Failure Modes in Benchmark-Validity Audits
• Abstract summary: We name five classes of pipeline failure and demonstrate each in a self-audit over safety benchmarks and open-weight instruction-tuned models.<n>We position the gate as a withholding and disclosure protocol for assurance-grade evidence, supplementary to (not a replacement for) classical construct-validity evidence.
• Score: 0.30586855806896046
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Governance frameworks ask AI providers and auditors for documented evaluation evidence, and perturbation-based construct-validity audits are a common form of that evidence. We argue the audits are themselves fragile: their conclusions can be silently manufactured by implementation details that readers cannot see in the reported numbers. We name five classes of pipeline failure and demonstrate each in a self-audit over safety benchmarks and open-weight instruction-tuned models. Under a unified six-point due-diligence gate, every cell lands in a non-confirmatory bucket, and no cell reaches confirmatory. The evidence here is a single two-model, five-benchmark case study, and F1--F5 is an illustrative, deliberately non-exhaustive starting taxonomy -- not a comprehensive partition of audit failures. We position the gate as a withholding and disclosure protocol for assurance-grade evidence, supplementary to (not a replacement for) classical construct-validity evidence, and not as a route to benchmark-validity verdicts.

Related papers

• VERA-8B: Evidence-Grounded Audit Risk Reasoning from SEC Filings [0.6445605125467574]
  We develop VERA-8B, a new end-to-end audit reasoning system.<n> VERA-8B identifies audit risks before enforcement actions occur.<n>We are the first to unify SFT and GRPO for evidence-grounded audit reasoning.
  arXiv  Detail & Related papers  (2026-08-28T14:55:49Z)
• IntelliAudit: Using Large Language Models to Evaluate Audit Controls [0.1759008116536278]
  We present IntelliAudit, a retrieval-grounded multi-agent system for IT audit evidence evaluation.<n>Given a control and an evidence corpus, IntelliAudit retrieves relevant artifacts, generates an evidence-grounded assessment.<n>We evaluate it across multiple simulated organizations using expert review and audit-readiness user feedback.
  arXiv  Detail & Related papers  (2026-08-07T18:21:06Z)
• From Trajectories to Evidence: Auditable Experimental Records for Industrial Research Agents [47.678562263387214]
  Research agents increasingly conduct multi-round machine-learning experiments in industrial recommendation settings.<n>Yet a completed trajectory is not automatically evidence: generated artifacts may be unsupported or incomplete, executed rounds may be invalid or confounded, and later modifications may obscure earlier findings.<n>We introduce an evidence-grounded framework that couples bounded verification of consequential artifacts with post-execution claim qualification.
  arXiv  Detail & Related papers  (2026-08-05T13:37:38Z)
• Coding Agents as Test-Suite Auditors: Finding What Official Suites Miss While Approaching What They Catch [47.836680625916266]
  An off-the-shelf coding agent, serving as a test-suite auditor, both builds adversarial test suites to expose what official suites miss.<n>A certification chain determines whether each agent-flagged submission is genuinely buggy without relying on the official judge.<n>On post-cutoff Codeforces problems with no available official suites, the same test-building method leads all five reproduced baselines at every tested input budget.
  arXiv  Detail & Related papers  (2026-08-03T05:24:03Z)
• White Box Evidence Packages for Policy Audit Reports [0.0]
  We show that internal evidence changes how reports cite and reason about evidence, but more internal citations do not by themselves make a report more valid.<n>This study reframes internal model access as an evidence design problem for audit, rather than as a guarantee of transparency.
  arXiv  Detail & Related papers  (2026-07-23T16:06:42Z)
• Reasoning Consistency Scanning: A Framework for Auditing Chain-of-Thought Validity in AI Safety Evaluations [0.0]
  We introduce reasoning consistency scanning, a reusable method for detecting inconsistency in chain-of-thought reasoning.<n>First, we formalize reasoning consistency as distinct from faithfulness and define a six-subtype taxonomy of inconsistency.<n>Third, we build a validated benchmark of 60 transcripts, manually adapted from InstrumentalEval outputs.<n>Fourth, we report results across four generator models and three evaluations from inspect_evals, showing that reasoning inconsistency is present, detectable, and varies systematically across both models and task types.
  arXiv  Detail & Related papers  (2026-07-08T10:13:19Z)
• Faults in Our Formal Benchmarking: Dataset Defects and Evaluation Failures in Lean Theorem Proving [13.87786274932945]
  Benchmarks for LLM-assisted theorem proving in Lean are often treated as intrinsically reliable because every solved instance comes with a machine-checked proof.<n>We audit five widely used Lean theorem-proving benchmarks and their forks, using corpus-scale static checkers to surface 4,833 findings.<n>We propose a fault taxonomy, a suite of automated checkers and recall-oriented semantic audit prompts, and release standards to guide the creation of formal math datasets.
  arXiv  Detail & Related papers  (2026-06-28T16:46:36Z)
• Failure Modes of Large Language Models on Research-Level Mathematics: A Taxonomy and an Empirical Characterisation [0.8594140167290097]
  The "First Proof" benchmark posed ten research-level mathematics questions to the strongest publicly available LLMs.<n>I identify four failure modes: citation fabrication (F1), premise smuggling (F2), silent problem reformulation (F3), and local-to-global compatibility gaps (F4)<n>I then audit eight one-shot proofs generated by Gemini 2.5 Flash on Questions 1, 2, and 5 of the benchmark, using two instruments built specifically to surface F1 and F2.
  arXiv  Detail & Related papers  (2026-06-12T14:58:20Z)
• Evidence Absence Is Not Evidence Insufficiency: Diagnosing NEI Construction Artifacts in Fact Verification [6.604874054866016]
  We introduce NEI-CAP, a construction-aware diagnostic protocol for insufficient-evidence evaluation.<n>We instantiate the protocol in SciFact-style scientific verification, with FEVER and HoVer as bounded external controls.
  arXiv  Detail & Related papers  (2026-05-26T07:51:33Z)
• Case-Grounded Evidence Verification: A Framework for Constructing Evidence-Sensitive Supervision [1.7044563989708406]
  We introduce case-grounded evidence verification.<n>We train a standard verifier on the resulting support task.<n>Results suggest that a major bottleneck in evidence grounding is not only model capacity, but the lack of supervision.
  arXiv  Detail & Related papers  (2026-04-10T17:55:38Z)
• DeepFact: Co-Evolving Benchmarks and Agents for Deep Research Factuality [49.62610727661819]
  Existing fact-checkers are primarily designed for general-domain, factoid-style atomic claims.<n>We propose Audit-then-Score (AtS), where benchmark labels and rationales are explicitly revisable.<n>We instantiate AtS as DeepFact-Bench, a versioned DRR factuality benchmark with auditable rationales.
  arXiv  Detail & Related papers  (2026-03-06T05:05:57Z)
• CausalT5K: Diagnosing and Informing Refusal for Trustworthy Causal Reasoning of Skepticism, Sycophancy, Detection-Correction, and Rung Collapse [1.4608214000864057]
  CausalT5K is a diagnostic benchmark of over 5,000 cases across 10 domains.<n>Unlike synthetic benchmarks, CausalT5K embeds causal traps in realistic narratives.<n>Preliminary experiments reveal a Four-Quadrant Control Landscape where static audit policies universally fail.
  arXiv  Detail & Related papers  (2026-02-09T17:36:56Z)
• OpenFactCheck: Building, Benchmarking Customized Fact-Checking Systems and Evaluating the Factuality of Claims and LLMs [59.836774258359945]
  OpenFactCheck is a framework for building customized automatic fact-checking systems.<n>It allows users to easily customize an automatic fact-checker and verify the factual correctness of documents and claims.<n>CheckerEVAL is a solution for gauging the reliability of automatic fact-checkers' verification results using human-annotated datasets.
  arXiv  Detail & Related papers  (2024-05-09T07:15:19Z)
• The Decisive Power of Indecision: Low-Variance Risk-Limiting Audits and Election Contestation via Marginal Mark Recording [51.82772358241505]
  Risk-limiting audits (RLAs) are techniques for verifying the outcomes of large elections. We define new families of audits that improve efficiency and offer advances in statistical power. New audits are enabled by revisiting the standard notion of a cast-vote record so that it can declare multiple possible mark interpretations.
  arXiv  Detail & Related papers  (2024-02-09T16:23:54Z)
• From Relevance to Utility: Evidence Retrieval with Feedback for Fact Verification [118.03466985807331]
  We argue that, rather than relevance, for FV we need to focus on the utility that a claim verifier derives from the retrieved evidence.<n>We introduce the feedback-based evidence retriever(FER) that optimize the evidence retrieval process by incorporating feedback from the claim verifier.
  arXiv  Detail & Related papers  (2023-10-18T02:59:38Z)
• QACHECK: A Demonstration System for Question-Guided Multi-Hop Fact-Checking [68.06355980166053]
  We propose the Question-guided Multi-hop Fact-Checking (QACHECK) system. It guides the model's reasoning process by asking a series of questions critical for verifying a claim. It provides the source of evidence supporting each question, fostering a transparent, explainable, and user-friendly fact-checking process.
  arXiv  Detail & Related papers  (2023-10-11T15:51:53Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.