Title: The Reliability Gap in Benchmark Auditing: Distribution Shift and Scale as Failure Modes of Contamination Detection

Abstract

The integrity of Large Language Model (LLM) assessments is compromised by benchmark contamination, a phenomenon where evaluation samples inadvertently enter the model’s training dataset. While statistical techniques for identifying training-data membership are available, their efficacy has been tested primarily within controlled academic environments characterized by large, uniform pre-training corpora and straightforward, single-stage training workflows. It remains uncertain whether these methods hold up under the complexities of real-world auditing.

This study highlights two significant, yet under-researched, failure modes: distribution shift, which occurs when the suspect and validation datasets do not adhere to the Independent and Identically Distributed (IID) assumption, and scale constraints, stemming from the fact that benchmarks are vastly smaller than pre-training corpora. We conducted a systematic evaluation of three prominent paradigms—LLM Dataset Inference, Post-Hoc Dataset Inference, and CoDeC. Our analysis covered 27 models across various families and sizes, ranging up to 27B parameters, including Pythia, OLMo~2, and specialized LLMs for cultural and medical domains. Furthermore, we extended this investigation to include frontier industry models.

Out of 335 total evaluations, only 199 produced accurate results. Specifically, LLM Dataset Inference generated false positives when distribution shift was present; Post-Hoc Dataset Inference lacked statistical power at the scale of benchmarks; and CoDeC offered only coarse provenance signals, which were inadequate for verifying individual benchmark splits. These findings expose a systematic reliability gap between controlled validation settings and practical benchmark auditing, demonstrating that statistical detection methods are not yet a viable substitute for transparent data provenance. To facilitate further research, we have open-sourced our benchmark.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC