Reproduce “Genesis: Visual Augmentation Primitives for Multimodal Chain-of-Thought Reasoning” 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
Fan Yang, Lifeng Chen, Yipu Wang, Han Wang, Feilong Chen, Yidong Ma, Zhiyang Chen, Chaoyang Zhao and 9 more
Genesis uses controlled image edits to create visual reasoning examples, then trains models to connect observations, reasoning, and answers. Its reward evaluates logical validity, completeness, and consistency alongside final answer correctness. On LogicVista, a 32B rubric-reward model scores 68.46, versus 66.22 with ordinary final-answer reinforcement learning, under matched training conditions. Genesis-32B scores 25.7% on UNREAL, versus 24.3%, a modest gain.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.