Leakscale measures how training exposure affects benchmark performance

Beyond Overlap: Estimating the Causal Effect of Benchmark Exposure

Computation and Language

Summary

Sometimes, test examples sneak into a model's training data, which can make it hard to tell if the model is genuinely smart or just remembering those examples. The authors created LeakScale, a way to measure how much this kind of 'cheating' improves a model's scores. It does this by designing new test tasks that use private information held back from training, then seeing how much better a model does when it has access to that private info. They found that exposure from training can boost performance by a noticeable amount across different models and tasks. This helps separate whether a model’s good score is because it saw similar examples before or because it actually learned the task well.

What this means in practice

  • For machine learning engineers: Quantify how much test data exposure inflates reported model accuracy to better interpret benchmark results.
  • For ai system auditors: Assess model robustness by measuring performance gaps due to private data exclusion in evaluation tasks.

Authors

Divyansh Singh

Abstract

Evidence that evaluation material entered training does not reveal how much it affected evaluation. This distinction leaves a contaminated benchmark score difficult to interpret: provenance can establish contact, but only a counterfactual can quantify the performance attributable to that contact. We present LeakScale, an interventional framework for estimating this missing quantity. LeakScale creates fresh executable tasks that require private, family-specific information absent from and non-derivable from the public task, controls access to that information, and estimates the resulting control-adjusted change in executable accuracy. Across 2,048 unique families, two model families, two executable domains, and 262,144 generations, exposure improves accuracy in every model-by-domain combination, with gains ranging from +7.17 to +27.31 percentage points. These findings separate two empirical questions that are often conflated: whether benchmark contact occurred and how strongly a reported score depends on it. LeakScale makes the latter directly measurable.
View PDFOpen arXiv