AI evidence extraction

Automated Benchmark Auditing for AI Agents and Large Language Models

Junlin Wang, Federico Bianchi, Shang Zhu, Fan Nie, Yongchan Kwon, Bhuwan Dhingra et al. · ArXiv.org · 2026

AI-generated evidence extraction, verified across multiple analytical personas. Not a substitute for the peer-reviewed original.

9/10

Relevance

2/4

Quality (LMQS)

Evidence

Citations

0.00

FWCI

Methodology & findings

Study design

Computational auditing framework using agentic inspection.

Main result

The study found that "ABA identifies critical issues including ambiguous task design, execution environment conflicts, and incorrect ground truths in over 25.7% of the evaluated tasks." Additionally, "filtering out these tasks with issues shifts model rankings and increases average performance on SWE-bench Verified and Terminal-Bench 2 by 9.9% and 9.6%, respectively."

Research paradigm

Empirical/Computational

Author conclusions

The authors conclude that "Modern benchmarks like SWE-bench and Terminal-Bench sit behind containerized environments, multi-stage evaluation harnesses, and grading logic that depends on runtime state" and that "Modern benchmarks now face a very similar condition" to complex software systems, suggesting that "We hope ABA can represent a step toward the analogous paradigm where an automated auditing framework can surface ambiguities, environment conflicts, and evaluation problems that domain experts may miss." They emphasize that "We release all findings and tooling publicly."

Risk of bias

Potential selection bias in benchmark choice (only benchmarks cited by at least 2 frontier model releases were included for frontier benchmarks); Agent model bias (Claude Opus 4.7 used as the auditing agent may have systematic blind spots); Manual review bias in validation (expert reviewers may have unconscious biases about benchmark issues); Sampling bias in trajectory mode (only recorded agent trajectories were audited, not all possible execution paths); Auditor bias: LLM-based auditor may have systematic blindspots or biases in issue detection; Selection bias: Benchmarks selected from frontier releases and NeurIPS may not represent all benchmark landscape; Validation bias: Ground truth validation relies on experts who may have authored the original benchmarks; Scope limitation: Exclusion of specialized domains may miss domain-specific issues; Selection bias: Benchmarks filtered by availability of verifiable ground truth, excluding subjective evaluation benchmarks; Agent configuration bias: Fixed to Claude Opus 4.7 default settings, which may not generalize to other LLM auditors; Expertise bias: Manual validation performed by domain experts who may share implicit assumptions with benchmark authors; Circular reasoning risk acknowledged for evaluation-methodology benchmarks; Coverage bias: Audit scope limited to 9 domains with exclusion of audio, video, embodied AI, remote sensing, and creative generation domains

Limitations

The authors note that "Ground truth is extremely scarce in this domain, and it is a non-trivial task to find a clean test set for this type of problem
The experts who could produce one are typically the same people who authored the benchmarks, and the issues we care about are precisely those they might have overlooked." Additionally, the audit scope is limited to benchmarks "where the system under consideration must be general-purpose LLMs or LLM agents" and excludes specialized domains or modalities such as audio, embodied AI, and remote sensing.

Open questions raised

Need for systematic automated auditing tools to catch issues that domain experts miss
Need for treating benchmarks as 'living artifacts' requiring continuous improvement rather than static evaluation instruments
Expansion of auditing frameworks to other specialized domains (audio, embodied AI, remote sensing) beyond the nine domains covered
Need for continuous improvement infrastructure in AI benchmarking analogous to software testing practices
Requirement for systematic benchmark auditing as part of standard development process
Need for standardized protocols to identify and fix benchmark issues at scale

Data: All audit findings and per-task evidence released as public artifact at https://www.autobenchaudit.com/; 168 audited benchmarks across nine domains; All task annotations and audit findings released as public artifact at https://www.autobenchaudit.com/; Per-task audit records, evidence bundles, and rubric annotations publicly released; 168 benchmarks across nine domains (34,285 tasks total) audited and findings releasedCode: https://github.com/IsThatYou/auto-bench-audit; Tool and all task annotations released to support future benchmark developmentExtracted from: pdfAgreement 45%

Explore related topics

Evaluation & Benchmarks Research Integrity Benchmarking