SWE-Review Paper: Agentic Code Review Loop Lifts Qwen3-30B Resolve Rate From 27.5% to 56.9% on SWE-Bench Verified, Achieves 6.5× Token Efficiency Over Independent Resampling
Summary
NTU + Huawei + HKU paper turns one-shot PR generation into a closed-loop generate-review-revise process. On SWE-Review-Bench (1,384 PRs from 500 SWE-bench Verified issues), the agentic review loop lifts Qwen3-30B-A3B resolve rate from 27.5% to 56.9% and Qwen3-Coder-30B-A3B from 50.9% to 68.8%. Releases the SWE-Review-Traj dataset of 8,914 decision-correct review trajectories.
Originally reported by huggingface.co
Read the original article →Original headline: SWE-Review Paper: Agentic Code Review Loop Lifts Qwen3-30B Resolve Rate From 27.5% to 56.9% on SWE-Bench Verified, Achieves 6.5× Token Efficiency Over Independent Resampling