Technical Summary: Contamination Detection via Context (CoDeC)

Problem Definition

The integrity of Large Language Model (LLM) evaluation is increasingly compromised by data contamination, where training data overlaps with benchmark datasets. This leads to inflated performance metrics driven by memorization rather than genuine generalization. Existing detection methods face significant limitations:

• Loss-based approaches (e.g., Vanilla Loss, Min-K%) often fail to distinguish between contaminated data and genuinely easy data, requiring extensive parameter tuning and external references.
• Explicit overlap checks require access to training corpora, which is rarely available for open-weight models.
• Membership Inference Attacks (MIAs) often struggle with scalability and interpretability, particularly for large models trained on massive corpora.

There is a critical need for an automated, scalable, and model-agnostic method to detect contamination at the dataset level without requiring access to training data or external reference models.

Methodology: Contamination Detection via Context (CoDeC)

CoDeC is a practical method that leverages In-Context Learning (ICL) to quantify contamination. It operates on the observation that LLMs react differently to in-context examples depending on whether the dataset was seen during training.

Core Hypothesis

• Unseen Datasets: When a model encounters a dataset it has not seen, adding in-context examples from that same dataset provides new information (style, structure, vocabulary), typically boosting the model's confidence (log-probabilities) on target tokens.
• Contaminated (Seen) Datasets: If the model has memorized the dataset, in-context examples provide no new information. Furthermore, these examples can disrupt memorization patterns, leading to a reduction in confidence compared to the baseline (no context).

The CoDeC Pipeline

The method requires only gray-box access (token probabilities) and follows four steps:

1. Baseline Prediction: For each datapoint $x$ in a suspect dataset $\mathcal{D}$, calculate the model's average log-likelihood on consecutive tokens.
2. In-Context Prediction: Sample $n$ additional examples from $\mathcal{D} \setminus \{x\}$, prepend them to $x$, and calculate the log-likelihood for $x$ in this new context.
3. Score Computation: Compute the difference $\Delta(x) = \text{logprob}_{\text{in-context}}(x) - \text{logprob}_{\text{baseline}}(x)$.
4. Aggregation: The contamination score $S_{\text{CoDeC}}(\mathcal{D})$ is the percentage of samples where $\Delta(x) < 0$ (i.e., confidence decreased with context).

$$S_{\text{CoDeC}}(\mathcal{D}) = \frac{1}{N} \sum_{i=1}^{N} \mathbb{1}[\Delta(x_i) < 0]$$

Key Properties

• Model-Agnostic: Works with any architecture (Transformers, RNNs, Mamba) and requires no retraining.
• Parameter-Free: No threshold tuning is required; the output is an intuitive percentage.
• Efficient: Requires only two forward passes per sample.
• Robustness: Detects contamination not only from exact matches but also from related distributions (e.g., synthetically generated data, paraphrases, or augmented content).

Key Contributions

1. Introduction of CoDeC: A simple, automated method for detecting and quantifying training data contamination using in-context learning dynamics.
2. Empirical Validation: Demonstration that CoDeC achieves near-perfect separation (AUC 99.9%) between seen and unseen datasets across diverse models (Pythia, GPT-Neo, RWKV, OLMo, Nemotron) and architectures, outperforming state-of-the-art baselines (Vanilla Loss, Min-K%, Zlib Ratio).
3. Analysis of Contamination Dynamics:
   • Training Phase: CoDeC scores rise sharply during the early stages of training (approx. 2% of training steps), indicating rapid memorization of distributional priors.
   • Fine-tuning: Fine-tuning on a dataset consistently drives CoDeC scores above 90%, confirming the method's sensitivity to recent exposure.
   • Contamination Transfer: The method detects contamination transfer between related benchmarks (e.g., MMLU to other QA datasets) and is robust against simple data augmentations (noise, whitespace changes).
4. Broad Evaluation: Application of CoDeC to over 40 recent open-weight models on widely used benchmarks (AIME, GSM8K, MMLU, etc.), revealing instances of potential contamination where models score close to or above the 80% threshold.

Results

• Separation Power: CoDeC achieves a dataset-level AUC of 99.9% across all evaluated models, whereas baselines show significant overlap between seen and unseen distributions (AUCs ranging from 64% to 92%).
• Score Interpretation:
  • >80%: Strong evidence of contamination (direct training or heavy reliance on memorized priors).
  • <20%: Typical for unseen datasets, indicating generalization.
  • 20%–60%: Ambiguous; may indicate partial contamination, related data exposure, or high dataset diversity acting as noise.
• Model Size Correlation: Larger models (e.g., 253B parameters) tend to exhibit lower contamination scores than smaller counterparts within the same family, suggesting that increased capacity facilitates generalization over memorization.
• Real-World Findings: The evaluation identified specific models with high contamination scores on benchmarks like AIME 2024 and RewardBench v1, suggesting potential data leakage or training on closely related synthetic data.

Significance and Claims

The paper positions CoDeC as a vital tool for restoring trust in LLM benchmarking. Its significance lies in:

• Practicality: Unlike previous methods, it requires no access to training corpora, external reference models, or complex tuning, making it immediately integrable into evaluation pipelines.
• Interpretability: The percentage-based score provides a clear, actionable metric for researchers and developers to assess the reliability of reported benchmark results.
• Detection of Subtle Contamination: By capturing the disruption of memorization patterns, CoDeC identifies contamination arising from related distributions and synthetic data, which traditional overlap checks often miss.

The authors conclude that while high CoDeC scores (>80%) indicate likely contamination, moderate scores should be interpreted via cross-model comparison to distinguish between dataset-specific effects and genuine memorization. The method supports the development of more reliable evaluation practices and helps identify models that rely on generalization rather than rote memorization.