Data Contamination or Genuine Generalization? Disentangling LLM Performance on Benchmarks

Large language models (LLMs) achieve high benchmark performance, but whether this stems from genuine generalization or data contamination remains unclear. This paper proposes a three-tier framework to disentangle these effects, combining n-gram alignment, canary insertion, and perturbation testing across open (Llama 2, Mistral) and closed (GPT-4) models. Our analysis reveals that: (1) Larger models exhibit higher contamination (18.1% n-gram overlap for Llama 2-70B) but smaller out-of-distribution (OOD) drops (−9.4%), suggesting scale mitigates memorization’s impact; (2) Perturbation experiments show factual recall is most vulnerable (e.g., −22.4% accuracy drop for entity swaps); (3) Chain-of-thought (CoT) evaluation uncovers hidden generalization (+21.3% gap for GPT-4), though humans outperform LLMs in robustness (e.g., −4.3% vs. −9.8% on swaps). We advocate for contamination-aware benchmarks and CoT-enhanced evaluation.