Search papers, labs, and topics across Lattice.
This paper critiques the conventional method of detecting contamination in LLM backtests, revealing that the standard comparison of scores before and after training cutoffs is ineffective. It demonstrates that structural factors lead to models appearing to have "leaked" knowledge due to their inherent understanding of recency, complicating the distinction between genuine skill and leakage. By introducing two novel measurement techniques鈥攃utoff identification and matched clean controls鈥攖he authors provide a way to adjust scores for leakage and validate their approach against ground truth, ultimately enhancing the reliability of backtesting in LLM evaluations.
Backtesting LLMs reveals that recency can mimic leakage, leading to misleading scores that obscure true model performance.
The standard check for contamination in LLM backtests is simple: compare scores before and after the training cutoff. We show this check is uninformative. Four flagship models fail it on questions they cannot have memorized: every scored question resolved after their cutoffs. The reason is structural. Models legitimately know more about times near their cutoff, so recency mimics leakage, and we prove no passive backtest can separate the two from genuine skill. Measurement, not just detection, requires information from outside the backtest. We supply it in two forms. A known cutoff identifies leakage at the boundary; a matched clean control identifies it globally and yields a leakage-adjusted score. We also derive where leakage hides: it concentrates on outcomes that surprised the crowd and were well covered in training, and partial memorization is disproportionately rewarded. We validate the estimators against ground truth by planting leakage in twin models, where they recover the injected dose and return null on clean questions. Deployed on frontier models, they detect one cutoff-localized signature and, at the audit's power floor, clear five models whose apparent advantages were recency alone. Backtests need not be discarded; they need one defensible reference.
