Reproduce “Thinking with Images as Continuous Actions: Numerical Visual Chain-of-Thought” 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

Kesen Zhao, Beier Zhu, Junbao Zhou, Xingyu Zhu, Zhongqi Yue, Hanwang Zhang

NV-CoT adds continuous coordinate outputs to visual chain-of-thought, letting a model specify image regions numerically, crop them, and use them. Rather than language tokens, box coordinates are treated as geometric values. On V* Bench, in a single-zoom comparison, DeepEyes-7B scored 86.4 and DeepEyes-7B + NV-CoT (RL) scored 89.0, a 2.6-point gain; the result does not establish continuity as the sole cause.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.