Groupwise Agentic Grading and Advantage Redistribution for Code Agent RL
The paper’s core move is to stop treating every passing code-agent rollout as equally useful training data.
GAGAR keeps rollout groups that contain both passing and failing trajectories, then has an SFT-trained agentic grader rank the passing candidates in a shared workspace. Lower-ranked passing solutions are downweighted, while the total advantage assigned to passing trajectories is preserved and shifted toward cleaner, more task-faithful implementations. The authors report industrial-scale evaluations on MiMo-V2.6-Flash and MiMo-V2.6-Pro checkpoints, with Flash code-only runs improving performance, reducing trajectory-length growth, and stabilizing training. HF Daily Papers' note
GAGAR keeps rollout groups that contain both passing and failing trajectories, then has an SFT-trained agentic grader rank the passing candidates in a shared workspace. Lower-ranked passing solutions are downweighted, while the total advantage assigned to passing trajectories is preserved and shifted toward cleaner, more task-faithful implementations. The authors report industrial-scale evaluations on MiMo-V2.6-Flash and MiMo-V2.6-Pro checkpoints, with Flash code-only runs improving performance, reducing trajectory-length growth, and stabilizing training. HF Daily Papers' note
score 5