SWE-Critix: Post-Training Generalist Verifiers for Assessing Coding-Agent Trajectories

SWE-Critix provides a post-training pipeline for training generalist verifiers to evaluate long-horizon coding-agent trajectories. Unlike scalar reward models, this method enables verifiers to judge and explain iterative interactions across hundreds of tool and repository steps. The approach addresses the lack of oracle evaluators in complex software engineering workflows.