Reproduce “VisionCoach: Reinforcing Grounded Video Reasoning via Visual-Perception Prompting” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Daeun Lee, Shoubin Yu, Yue Zhang, Mohit Bansal
VisionCoach trains a video reasoner with question-conditioned visual cues on hard examples, then distills higher-reward prompted trajectories so inference can use raw video without an added visual tool. Its identity-aware spatial reward also checks object names, timestamps, and boxes. On V-STAR, it reports 34.3 mAM and 47.5 mLGM, versus 19.3 and 22.4 for Qwen2.5-VL-7B, under the reported setup.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.