VeriFine: Scaling Verification for Self-Improvement in Embodied Reasoning
VeriFine makes the verifier part of the self-improvement loop, not a fixed judge.
The paper proposes an agent harness where the policy, curriculum, and judge co-evolve as new failures appear. A rubric judge identifies recurring errors, builds adaptive training data, and guides policy optimization. When the judge becomes the bottleneck, the system queries humans on selected failure cases and recalibrates the judge through human-agent disagreement resolution. The authors report continued gains on driving and robot navigation tasks under both reinforcement learning and supervised fine-tuning. ArXiv · AI/CL/LG's note
The paper proposes an agent harness where the policy, curriculum, and judge co-evolve as new failures appear. A rubric judge identifies recurring errors, builds adaptive training data, and guides policy optimization. When the judge becomes the bottleneck, the system queries humans on selected failure cases and recalibrates the judge through human-agent disagreement resolution. The authors report continued gains on driving and robot navigation tasks under both reinforcement learning and supervised fine-tuning. ArXiv · AI/CL/LG's note
score 5