Provable Joint Decontamination for Benchmarking Multiple Large Language Models

2026-05-23 · Source: cs.LG updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, extended

Summary

Benchmark data contamination poses a significant challenge in evaluating Large Language Models (LLMs), as evaluation examples appearing in training data can inflate performance and undermine cross-model comparisons. While existing score-based and single-model conformal methods lack theoretical guarantees or produce model-specific benchmarks, this work introduces Joint Envelope Conformal Selection (JECS). JECS formalizes multi-model decontamination as a joint selection problem, providing a conformal procedure that controls the Global Contamination Rate (GCR) across multiple audited models. Experiments on WikiMIA and ArXivTection, using models like GPT-NeoX-20B, Pythia-6.9B, and LLaMA-7B, demonstrate that JECS consistently maintains target GCR control while achieving higher power than the Joint Max Conformal Selection (JMCS) baseline, for instance, improving power from 0.094 to 0.447 on ArXivTection with Pythia-6.9B at α=0.1.

Key takeaway

For MLOps Engineers and AI Scientists evaluating multiple LLMs, benchmark contamination can severely skew results. You should adopt JECS to construct shared, decontaminated benchmarks, ensuring fair and reliable cross-model comparisons with provable Global Contamination Rate (GCR) control. This method significantly improves selection power over naive approaches, allowing you to identify more truly pure samples for robust evaluation.

Key insights

JECS provides provable global contamination rate control for shared LLM benchmarks, improving power over baselines.

Principles

• Benchmark contamination requires joint selection for fair multi-model comparison.
• Max-p aggregation of conformal p-values is valid but conservative.
• Envelope fitting can recalibrate p-values to improve selection power.

Method

JECS computes per-model conformal p-values, aggregates them by maximum, reconstructs a conservative envelope of the max-p null distribution, and applies an adaptive Benjamini–Hochberg (BH) procedure to envelope-rescaled values.

In practice

• Use JECS to create shared, decontaminated benchmarks for multiple LLMs.
• Apply JECS with various detection scores (e.g., Min-K%++, Perplexity).
• Ensure a shared calibration set of known members for all audited models.

Topics

• LLM Evaluation
• Benchmark Contamination
• Conformal Prediction
• False Discovery Rate
• Membership Inference
• Statistical Guarantees

Code references

• lyy1994/awesome-data-contamination

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, MLOps Engineer

Related on AIssential

• CoEval: Ranking Language Models for Custom Tasks Without Labeled Data or Trustworthy Benchmarks — The CoEval framework enables reliable, label-free LLM ranking for custom tasks by synthesizing fresh benchmarks and using diverse judge ensembles.
• Neural Bayesian Anomaly Mitigation: A Robust Loss that Doubles as an Unsupervised Contamination Classifier — NBAM is a robust loss that simultaneously classifies contamination using a Bayesian latent-switch mixture model.
• Crossing the Validation Crisis: Cross-Validation Reduces Benchmarking Variance Surprisingly Well — Cross-validation significantly reduces benchmarking variance, improving reliability of machine learning performance evaluation through virtual data augmentation.
• The Reliability Gap in Benchmark Auditing: Distribution Shift and Scale as Failure Modes of Contamination Detection — Current statistical methods for detecting LLM benchmark contamination are unreliable in realistic auditing due to distribution shift and scale.
• Search-Time Contamination in Deep Research Agents: Measuring Performance Inflation in Public Benchmark Evaluation — Deep research agents' web search can contaminate benchmark evaluations, inflating performance by bypassing genuine reasoning.
• Did the Model See the Benchmark During Training? Detecting LLM Contamination — A lightweight method estimates LLM benchmark contamination by analyzing model behavior when training data is unknown.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.LG updates on arXiv.org.