BenchChecker improves trust in bug-fixing tests for code AI models
BenchChecker: Assessing the Credibility of Bug-Fixing Benchmarks for LLMs
Software Engineering
Summary
Bug-fixing tests used to check how well AI models fix code often have a problem: some test problems are accidentally included in the AI's training data, making the AI look better than it really is. The authors created BenchChecker, a method that uses clues from code history and the AI’s own answers to spot these overlaps. Their tests show BenchChecker works well and that removing these overlaps gives a more honest view of AI model performance. This helps people better understand which AI really fix bugs well.
What this means in practice
- •For software development teams: Use BenchChecker to confirm bug-fixing AI tools are assessed fairly by filtering out test cases included in AI training data.
- •For software quality assurance teams: Apply BenchChecker to verify benchmarks that evaluate code repair tools to trust reported bug resolution rates.
Authors
Di Wu, Xu He, Shu Wang, Kun Sun
Abstract
While Large Language Models (LLMs) show promise for Automatic Program Repair (APR), existing benchmarks suffer from data contamination, in which test samples appear in training sets, artificially inflating performance. To address this, we propose BenchChecker, a framework compatible with both open-source and closed-source models that supports more credible evaluation by detecting such contamination. Grounded in the observation that persistent patches likely remain in recent repository snapshots used for training, BenchChecker employs a forensics-based two-stage detection process using only textual model outputs together with public repository history. The first stage conducts a repository presence test by probing the model with code prefixes to infer training inclusion; the second stage performs a patch presence test to verify patch persistence via similarity alignment and test-suite verification. Evaluations using StarCoder-15B, StarCoder2-15B, and CodeGen2-16B trained on The Stack v2 demonstrate that BenchChecker achieves robust detection with AUC scores around 0.9 across six benchmarks. Furthermore, applying BenchChecker to SWE-bench Verified and Multi-SWE-bench reveals that filtering contaminated samples reduces the reported resolution rates of most LLMs by more than 20% relative to the original rates on medium-difficulty tasks.