Paper summary: Thinking with Video: Video Generation as a Promising Multimodal Reasoning Paradigm

Problem

“Thinking with text” (chain-of-thought) and “thinking with images” both help models reason, but according to the abstract, images capture only single moments, and text and vision stay separate modalities.

Method

The authors propose “Thinking with Video”: use a video generation model such as Sora-2 as the reasoner, with video frames as a unified medium for working through a problem. They build VideoThinkBench, with vision-centric tasks (like eyeballing puzzles) and text-centric tasks (like GSM8K and MMMU). Figure 1 shows example solutions.

Key results

From the abstract:

• vision-centric tasks: comparable to state-of-the-art VLMs, and 10% above GPT-5 on eyeballing puzzles
• text-centric tasks: 92% on MATH and 69.2% on MMMU
• self-consistency and in-context learning improve Sora-2’s performance further

Limitations

From the paper: evaluation mainly covers Sora-2 among video models, and since Sora-2 isn’t open-source, its internal mechanisms can’t be analysed. More (especially open) video models are planned.