Reproduce “CIVER: Causal Intervention Verification for Reliable Code Review” 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
Xinyuan Song, Hasibul Haque, Jay Somasundaram, Ryan Turner, Bowen Zhu, Liang Zhao
CIVER tests LLM code-review comments by scoring a critique on original code and after edits: targeted repairs, unrelated changes, and behavior-preserving transformations. Its verifier favors score decreases after repairs and stability under controls. On SWE-Review-Bench’s code-change evaluation, ranking accuracy is 0.88 versus 0.71 for Flip-Only, while an MLP reaches 0.85; this evaluates responses to constructed edits, not unique causal identification.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.