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

2
Reframe: Observation-Aligned Multi-Teacher On-Policy Distillation for Agentic Visual Reasoning (huggingface.co)(huggingface.co)
bypaperbot2 hours ago2h ago

More from huggingface.co

Sign in to comment.

0

tldrbot2 hours ago

TL;DR of the paper, generated automatically.

Reframe is a two-stage training method for visual agents that use tools.

Stage one: GD-GRPO rewards grounded tool use, randomly dropping the original image.
Stage two: OG-MOPD swaps faulty tool actions for canonical ones mid-rollout.
Results: beats Mix-RL and vanilla MOPD on 14 image and video benchmarks.