Benchmark Contamination Detectability: Information Limits and Power-Calibrated Audits
A recent study titled 'When Is Benchmark Contamination Detectable? Information Limits and Power-Calibrated Audits' explores how to identify contamination in benchmark datasets. It differentiates between clean benchmarks and low-power audits, treating benchmarks as a sparse combination of clean and previously observed controls. The ability to detect contamination is influenced by the fraction of contamination, the separability of behaviors, and the size of the sample. The research introduces the concept of detector efficacy, derived from control data, along with a sample-split certificate that establishes a minimum threshold for contamination. Findings indicate that frozen calibration efficacy accurately predicts power curves with high R-squared values, although Gaussian budgets tend to miscalibrate with smaller samples. A two-stage planner corrects budgets conservatively. The research underscores the importance of audit interpretation in addition to efficacy and validity, making it a vital contribution to the AI community.
Key facts
- The paper formalizes the distinction between a clean benchmark and an audit with low power.
- Detectability is governed by alpha * rho * sqrt(m), where rho^2 = chi^2(P_1 || P_0).
- Any scalar detector reduces to its efficacy, ef = |E_1 f - E_0 f| / sqrt(Var_0(f)) <= rho.
- Efficacy can be estimated from controls before the audit is run.
- A sample-split certificate lower-bounds alpha distribution-free.
- Frozen calibration efficacy predicts held-out power curves with R^2 = 0.83-0.98 across six exact-permutation channels.
- The efficacy-only Gaussian budget fails in 9/9 gate-passing channels.
- A predeclared two-stage planner repairs the budgets and is uniformly conservative.
- The certificate is valid but vacuous at audit scale.
- A five-seed paired injection study recovers the mechanism ordering verbatim > paraphrase > surface.
Entities
Institutions
- arXiv