The methods researchers use to detect benchmark contamination in large language models are less reliable than the field assumes.
A paper published on arXiv evaluated three leading statistical detection methods — LLM Dataset Inference, Post-Hoc Dataset Inference, and CoDeC — across 335 experiments covering 25 models from families including Pythia and OLMo 2, at scales up to 27 billion parameters. Only 201 of those 335 evaluations produced correct outcomes. The researchers identified two root causes: distribution shift, which breaks the assumption that suspect and validation data come from the same statistical population, and scale constraints, because benchmarks are orders of magnitude smaller than the pre-training corpora the tools are designed to analyze. They also extended the analysis to frontier commercial models.
This matters because benchmark contamination — when evaluation examples leak into training data — is the central threat to trusting AI leaderboard scores. If the tools meant to catch cheating are themselves unreliable, the entire premise of independent auditing is shakier than it looks. The study's finding that none of the three methods works cleanly in realistic conditions undercuts claims that contamination is a solved or solvable problem with current statistical techniques.
The honest takeaway is that transparent data provenance — actually documenting what went into training — remains the only credible alternative, which is precisely what most frontier labs have resisted disclosing.