Reframe: Observation-Aligned Multi-Teacher On-Policy Distillation for Agentic Visual Reasoning
Reframe repairs weak tool actions so visual agents can learn from better evidence.
The framework addresses two failures: agents overlooking tool-returned observations and poorly grounded actions leaving too little evidence for training. Its first stage trains spatial and temporal specialists with evidence-aware rewards, sometimes removing the original visual input to encourage use of tool results. Its second stage replaces deficient tool actions with canonical ones, then resumes distillation from the repaired observation. The authors report better results than Mix-RL and standard multi-teacher on-policy distillation across 14 image and video benchmarks, without a task-specific supervised fine-tuning cold start.
HF Daily Papers' note
The framework addresses two failures: agents overlooking tool-returned observations and poorly grounded actions leaving too little evidence for training. Its first stage trains spatial and temporal specialists with evidence-aware rewards, sometimes removing the original visual input to encourage use of tool results. Its second stage replaces deficient tool actions with canonical ones, then resumes distillation from the repaired observation. The authors report better results than Mix-RL and standard multi-teacher on-policy distillation across 14 image and video benchmarks, without a task-specific supervised fine-tuning cold start.
HF Daily Papers' note
score 5