Quantifying Memorization Advantage in Code LLMs

---

Résumé

The lack of transparency regarding the code datasets used during LLM training creates substantial challenges in detecting, evaluating, and mitigating data leakage. This paper applies a perturbation-based approach to quantify the “memorization advantage” of LLMs across various coding tasks by measuring the performance gap between a model’s handling of data it has likely encountered during training versus novel inputs. Our comprehensive analysis examines 8 open-source code LLMs across 19 benchmark datasets spanning four distinct categories: standard code generation, code understanding, security vulnerability detection, and bug identification. The results reveal significant variations in sensitivity patterns, with models like StarCoder exhibiting substantially higher sensitivity scores (up to 0.8) on certain benchmarks like APPS compared to models like QwenCoder, which maintained consistently lower values (< 0.4) across most benchmarks, suggesting fundamental differences in their generalization process and their learned knowledge. Different task categories also showed distinct patterns, with code summarization demonstrating low sensitivity (< 0.3) and test generation tasks exhibiting significantly higher values (0.4-0.7, p < 0.001).
Interestingly, our analysis of the widely-used CVEFixes and Defects4J benchmarks, frequently suspected of data leakage in the research community, reveals unexpectedly low memorization advantage scores across all models. Defects4J demonstrated significantly lower sensitivity (0.2-0.4, p < 0.01) compared to other program repair benchmarks (0.5-0.8), while CVEFixes showed consistently low values below 0.1. These findings challenge prevailing concerns about these datasets’ validity for evaluating code LLMs and suggest that models may be effectively generalizing from this data rather than merely memorizing it. Our findings provide critical insights into the generalization capabilities of code LLMs and emphasize the need for more robust evaluation frameworks, particularly in security-related domains where the widest range of sensitivity distributions (from < 0.1 to > 0.8) indicates variable generalization challenges.

Mots-clés

frameworks, particularly i