Reinforcement Learning with Visually Attributable On-Policy Self-Distillation
A visual-evidence filter beat GRPO and on-policy self-distillation across six visual reasoning benchmarks, the authors report.
Teacher corrections can reflect language priors or response style rather than what an image shows, potentially reinforcing errors. The proposed Visual-Attribution Gating compares teacher views with visual evidence present and degraded, accepting only corrections whose direction that contrast supports. It combines those filtered token-level corrections with task-level verifiable rewards. The authors report state-of-the-art average performance, though the abstract gives no numerical gains.
ArXiv · AI/CL/LG's note
Teacher corrections can reflect language priors or response style rather than what an image shows, potentially reinforcing errors. The proposed Visual-Attribution Gating compares teacher views with visual evidence present and degraded, accepting only corrections whose direction that contrast supports. It combines those filtered token-level corrections with task-level verifiable rewards. The authors report state-of-the-art average performance, though the abstract gives no numerical gains.
ArXiv · AI/CL/LG's note
score 4