English

Detecting Benchmark Contamination Through Watermarking

Cryptography and Security 2025-07-22 v2 Artificial Intelligence

Abstract

Benchmark contamination poses a significant challenge to the reliability of Large Language Models (LLMs) evaluations, as it is difficult to assert whether a model has been trained on a test set. We introduce a solution to this problem by watermarking benchmarks before their release. The embedding involves reformulating the original questions with a watermarked LLM, in a way that does not alter the benchmark utility. During evaluation, we can detect ``radioactivity'', \ie traces that the text watermarks leave in the model during training, using a theoretically grounded statistical test. We test our method by pre-training 1B models from scratch on 10B tokens with controlled benchmark contamination, and validate its effectiveness in detecting contamination on ARC-Easy, ARC-Challenge, and MMLU. Results show similar benchmark utility post-watermarking and successful contamination detection when models are contaminated enough to enhance performance, \eg pp-val =103=10^{-3} for +5%\% on ARC-Easy.

Keywords

Cite

@article{arxiv.2502.17259,
  title  = {Detecting Benchmark Contamination Through Watermarking},
  author = {Tom Sander and Pierre Fernandez and Saeed Mahloujifar and Alain Durmus and Chuan Guo},
  journal= {arXiv preprint arXiv:2502.17259},
  year   = {2025}
}
R2 v1 2026-06-28T21:55:41.119Z