SWE-Critix trains agent critics for code repair

Researchers affiliated with Huawei and Monash University Malaysia presented SWE-Critix, a training method for verifiers that judge whether a coding agent’s execution trajectory completed its task. A teacher model compares successful and failed trajectories for the same issue to produce training explanations, followed by supervised fine-tuning and reinforcement learning with rewards based on test outcomes.

The final system reached 78% overall accuracy in the paper’s CodeForge Preview evaluation. The authors measured this on a test set derived from the same data source; they say test-reward training helped distinguish successful repairs, while recall remained low for some error classes. This early benchmark result does not establish repair success on real repositories.