Auditing VLM Benchmarks from Existing Evaluation Runs
We reuse those artifacts to audit the benchmark itself, and add an audit layer to VLMEvalKit: six detectors that run as post-processing over saved outputs, plus one additional text-only rerun with visual inputs removed, emitting a per-question report.
ProofPaper ↗
Key points
- Every evaluation run of a vision-language model (VLM) already produces per-question artifacts (predictions, judged correctness, metadata) that leaderboards reduce to a single score and then discard.
- The layer reports per-question findings together with aggregate audit statistics; for the headline quantities, it attaches permutation-null excess where applicable, interval estimates for proportions, and sensitivity to the model pool.
- Auditing MMStar, a benchmark curated specifically so that the image is indispensable, with three models, we find zero aggregate accuracy gain from the image on 20.2% of its questions.
- This excess is concentrated in the mathematics and science categories and is absent from coarse perception, a localization consistent with knowledge substituting for visual evidence; confirming that mechanism would require item-level review.
Sources (1)
- [1]Auditing VLM Benchmarks from Existing Evaluation RunsarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 9, 01:05 PM
We reuse those artifacts to audit the benchmark itself, and add an audit layer to VLMEvalKit: six detectors that run as post-processing over saved outputs, plus one additional text-only rerun with visual inputs removed, emitting a per-question report.
Every evaluation run of a vision-language model (VLM) already produces per-question artifacts (predictions, judged correctness, metadata) that leaderboards reduce to a single score and then discard.
Extractive summary: sentences quoted from the sources.