Reproduce “AICD Bench: A Challenging Benchmark for AI-Generated Code Detection” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Daniil Orel, Dilshod Azizov, Indraneil Paul, Yuxia Wang, Iryna Gurevych, Preslav Nakov

AICD Bench introduces the most comprehensive benchmark for AI-generated code detection, comprising 2 million code samples across 9 languages and 77 large language models. Extensive evaluations reveal that current detection methods largely fail to generalize to unseen code distributions, highlighting the inadequacy of existing techniques for real-world application.

Key insights

  • Current AI-generated code detection methods exhibit poor generalization and robustness, with performance significantly degrading when faced with unseen programming languages, application domains, or generator families.
  • Domain shift presents a greater challenge to detection performance than programming language shift.
  • Hybrid (human-LLM co-authored) and adversarial code generation strategies effectively bypass existing detectors or are frequently misclassified, indicating their practical effectiveness.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.