Recoleta
GitHub
Source note

BenchGuard: Who Guards the Benchmarks? Automated Auditing of LLM Agent Benchmarks

Authors
Xinming Tu, Tianze Wang, Yingzhou, Lu, Kexin Huang, Yuanhao Qu, et al.
Source
arXiv
Published
Collection
Software Intelligence
LLM Agent BenchmarksBenchmark AuditingExecution Based EvaluationHuman AI ReviewCode Evaluation
Open arXivMarkdown source

Summary

BenchGuard audits execution-based LLM agent benchmarks by checking whether instructions, reference solutions, evaluation scripts, and environments agree. The paper claims it finds benchmark defects that human reviewers missed, at low audit cost.

Problem

  • Execution-based benchmarks can score agents incorrectly when task text, gold code, evaluator logic, or environment settings conflict.
  • These errors matter because they can make valid agent solutions fail, make tasks unsolvable, and distort leaderboard scores.

Approach

  • BenchGuard takes four artifacts per task: instruction, ground-truth program, evaluation script, and environment configuration.
  • A structured LLM audit checks the artifacts in six steps: task understanding, gold-program correctness, evaluator logic, task specification, environment checks, and deduplication.
  • Deterministic static checks run with the LLM audit, and optional agent programs or execution logs add evidence when available.
  • Findings are labeled by a 4-category, 14-subcategory defect taxonomy covering ground truth, evaluation, instruction, and environment issues; each finding includes severity, confidence, and cited evidence.

Results

  • On ScienceAgentBench, the audit found 12 defects confirmed by the original authors across 102 tasks, including fatal task errors, metric mismatches, and evaluators that reject correct outputs.
  • On ScienceAgentBench definition-only auditing, Claude Opus 4.6 reached 83.3% exact recall and 91.7% recall with partial matches; the five-model union reached 91.7% exact recall and 100% with partial matches.
  • Adding agent programs on ScienceAgentBench raised Claude Opus 4.6 to 91.7% exact recall and 100% recall with partial matches; the five-model union reached 100% exact recall.
  • On BIXBench Verified-50, the five-model union exactly matched 20 of 24 expert-identified atomic issues, 83.3% exact recall, and matched 23 of 24 with partial matches, 95.8% recall.
  • The best single BIXBench model, Claude Opus 4.6, exactly matched 13 of 24 issues, 54.2% recall, at $5.98 for 50 tasks.
  • Full five-model audits cost $14.38 for 50 BIXBench tasks and $22.72 for 102 ScienceAgentBench tasks in definition-only mode; the paper also states that a 50-task bioinformatics audit costs under $15.

Link