벤치마크 오염

Large language models are trained on web-scale text that includes GitHub, arXiv papers, forums and tutorial sites. Popular benchmarks are public by design, so their questions, answers and discussions of them get copied across the web and swept into training data. Contamination can be direct, where the exact test item appears, or indirect, through paraphrases, translations, solution write-ups, or synthetic data generated by another model that had seen the benchmark. It can also enter during fine-tuning if instruction datasets were assembled from benchmark-like sources. A useful distinction is between seeing only the question and seeing the question with its answer. The second is more damaging, because the model can recall the answer instead of working it out. The result is an inflated score that does not transfer to new problems. Lab awareness is not new. The GPT-3 paper (2020) ran a 13-gram overlap analysis between its benchmarks and training data and acknowledged that a bug left some overlaps in place. When training data is available, overlap search is the most direct check. When it is not, researchers use indirect methods: prompting the model with the start of a test item to see if it completes the rest verbatim; membership inference methods such as Min-K% Prob (Shi and colleagues, 2023), which look at how confidently a model predicts a text's tokens; comparing scores on original items against rewritten or newly written equivalents; and time-based splits that compare problems created before and after the training cutoff. Two misconceptions are common. First, overlap does not always raise scores much; some studies find modest effects on certain tasks. Second, finding no exact overlap does not prove a benchmark is clean, because paraphrased or translated copies slip past string matching. Mitigations include private held-out test sets, canary strings, regularly refreshed benchmarks such as LiveBench and LiveCodeBench, decontamination filters, and publishing contamination analyses alongside results.