Overview
Benchmark data contamination presents a challenge in Large Language Model (LLM) evaluation. This occurs when evaluation examples are present in the training data of one or more audited models, potentially inflating reported performance and affecting the reliability of cross-model comparisons. The presented work addresses this by formalizing multi-model benchmark decontamination as a joint selection problem.
It introduces Joint Envelope Conformal Selection (JECS), a conformal procedure designed to enable global contamination rate (GCR) control under specified assumptions. JECS computes per-model conformal p-values, aggregates them using the per-item maximum, and reconstructs a conservative envelope of the max-p null distribution from right-tail observations above a data-driven threshold. The adaptive Benjamini-Hochberg (BH) procedure is then applied to the envelope-rescaled values to select a benchmark with provable GCR control. Empirical evaluations across various models and benchmarks indicate that JECS achieves higher power compared to a max-p baseline, while consistently maintaining the target GCR control.
Research Context
The evaluation of LLMs faces a central challenge due to benchmark data contamination. This issue arises when evaluation examples are included in the training data of models being assessed. Such inclusion can lead to inflated performance metrics for the affected models and can compromise the integrity of comparisons across different models.
Existing approaches for detecting training data typically design scores to quantify the degree to which a model has memorized a given data point. However, these score-based methods lack theoretical guarantees regarding their effectiveness. More recent initiatives have utilized conformal methods to achieve provable false-identification control for individual models. A limitation of applying these conformal approaches separately to each model is that it can result in model-specific benchmarks, which may undermine the objective of fair comparison across multiple models.
Approach
The research formalizes the challenge of multi-model benchmark decontamination as a joint selection problem. To address this, it proposes Joint Envelope Conformal Selection (JECS), a conformal procedure designed to provide global contamination rate (GCR) control under stated assumptions. The JECS procedure involves several steps:
- Per-model conformal p-value computation: For each model, conformal p-values are calculated.
- Aggregation by per-item maximum: The computed p-values are then aggregated by taking the maximum value for each item across all models.
- Envelope reconstruction: A conservative envelope of the max-p null distribution is reconstructed. This reconstruction is based on right-tail observations that exceed a data-driven threshold.
- Adaptive Benjamini-Hochberg (BH) procedure application: The adaptive BH procedure is applied to the values that have been rescaled by the reconstructed envelope.
- Benchmark selection: This application of the BH procedure enables the selection of a benchmark with provable GCR control.
Findings
The proposed Joint Envelope Conformal Selection (JECS) procedure enables global contamination rate (GCR) control under specified assumptions. Extensive experiments were conducted across various models and benchmarks to evaluate its performance. These experiments demonstrated that JECS achieves higher power compared to the max-p baseline. Furthermore, JECS consistently maintained the target GCR control throughout these evaluations.
Why This Matters
The reliability of LLM evaluations and comparisons is compromised by benchmark data contamination. A method that offers provable control over contamination rates can enhance the trustworthiness of reported model performances and foster more equitable cross-model assessments. The ability to control the global contamination rate can assist in developing more robust and fair evaluation frameworks for LLMs.