Haozhi Yuan, Handong Li, Yifan Zhang, Qianshan Wei (Nanjing University) · Visual Agent Distillation

Reframe: Observation-Aligned Multi-Teacher On-Policy Distillation for Agentic Visual Reasoning

Reframe is a post-training framework for visual agents that aligns the observation state returned by tools, not only the output distribution. It combines Ground-Drop-GRPO, which trains spatial and temporal specialists with evidence-aware rewards and stochastic removal of the input image, and Observation-Gated-MOPD, which replaces deficient tool actions mid-rollout and resumes distillation. The authors report that Reframe outperforms Mix-RL and vanilla multi-teacher on-policy distillation across 14 image and video benchmarks, and that no task-specific SFT cold start is needed.