Reproduce “FrameGRPO: Incentivizing Video Reasoning via Multi-Frame Rollouts and Temporal Credit” 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
Jiaxuan Sun, Biao Wang, Shuai Li
FrameGRPO trains video-reasoning models by generating answers to the same question from temporally aligned views with different frame densities, then selectively crediting correct sparse-view answers when full-view attempts fail. On VSI-Bench, Qwen2.5-VL-7B accuracy rises from 34.1% with GRPO to 37.9%; Qwen2-VL-7B slightly declines there, and sparse views may miss decisive events.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.