NEW YORK — What happened: A consortium of independent machine learning researchers published a comprehensive audit showing that nearly 70% of public evaluation benchmarks have significant overlap with the web corpora used to train modern frontier models, violating core scientific methodology.
Why it matters: Goodhart's law states that when a measure becomes a target, it ceases to be a good measure. If labs are inadvertently or intentionally training models on evaluation test sets, reported capabilities are inflated, leading to misinformed enterprise deployment and misplaced safety trust.
Deep dive: The audit utilized advanced n-gram and semantic embedding distance checks to scan the training sets of twelve major commercial and open-weights models against standard benchmarks like MMLU-Pro and GPQA. The findings revealed that even when exact test items were excluded, paraphrased versions were frequently embedded deep within pre-training tokens.
Report check (claims vs what is verified vs still rumor): The audit paper claims that benchmark scores are inflated by an average of 14 percentage points due to contamination. This metric has been verified by replication studies at two separate academic institutions. Claims that specific labs intentionally poisoned datasets to game leaderboards remain unproven and are categorized as speculation.
Open questions: Can the industry move toward private, dynamically generated, or interactive evaluation environments before public leaderboards become entirely obsolete?