SWE-Review: Closing the Loop on Issue Resolution with Agentic Code Review

SWE-Review from NTU, Huawei Technologies, and HKU reframes AI code review as a closed-loop mechanism for issue resolution rather than a post-hoc comment generator. The paper builds SWE-Review-Bench with 1,384 AI-generated pull requests from 500 SWE-bench Verified issues and SWE-Review-Traj with 8,914 decision-correct review trajectories, then evaluates agentic reviewers such as Claude Opus 4.6, GLM-5, and Qwen3 variants. Compared with single-turn fixed-context review, agentic review improves both decision accuracy and downstream resolve rate, especially on non-local bugs: for Qwen3-30B-A3B, resolve rate after revision rises from 27.5% to 52.6%, while the generate-review-revise loop lifts Qwen3-30B-A3B from 27.5% to 56.9% and Qwen3-Coder-30B-A3B from 50.9% to 68.8%. The strongest reviewer, Claude Opus 4.6, reaches 89.4% decision accuracy on one split and 81.8% weighted-average accuracy with only 148K tokens on average, and review-guided iterative revision outperforms verifier-selected best-of-N by reaching 38.4% resolve rate from a 22.9% baseline using only 2.44 samples on average. The key technical claim is that structured, repository-exploring review feedback is not just a gate, but a reusable supervision signal for training, revision, and test-time scaling.