Reframe Aligns Observations in Tool-Using Visual Agents
A paper on arXiv introduces Reframe, a post-training framework that aligns observation states as well as output distributions for agentic visual reasoning. Its Ground-Drop-GRPO stage trains spatial and temporal specialists with evidence-aware rewards and stochastic removal of original visual inputs. Observation-Gated-MOPD then replaces deficient tool actions with canonical ones during rollouts and resumes distillation from the repaired observations. The authors report that Reframe outperformed Mix-RL and vanilla multi-teacher on-policy distillation across 14 image and video benchmarks; ablations showed that GD-GRPO increased observation dependence and tool gain compared with outcome-based GRPO.
Reframe combines observation-aligned reinforcement learning with observation-repaired distillation to address visual agents that may overlook tool-returned evidence or take poorly grounded actions. The authors report gains across 14 image and video benchmarks, although the results have not been independently verified in the supplied report.
Computer Science > Computer Vision and Pattern Recognition
arXiv:2611.12886 (cs)
[Submitted on 9 Nov 2026]
Title:Reframe: Observation-Aligned Multi-Teacher On-Policy Distillation for Agentic Visual Reasoning
Authors:Haozhi Yuan, Handong Li, Yifan Zhang, Qianshan Wei
View a PDF of the paper titled Reframe: Observation-Aligned Multi-Teacher On-Policy Distillation for Agentic Visual Reasoning, by Haozhi Yuan and 3 other authors