arXiv 2511.04570 introduces "Thinking with Video," a paradigm where the LLM uses video as the reasoning medium, rather than text. The standout: Sora-2 fine-tuned for "video reasoning" hits 92% on the MATH benchmark — significantly above text-based reasoning models (GPT-5.6 at 87.4%, Claude Opus 4.7 at 88.1%).
The "video as reasoning" insight: text is a lossy representation of thought — many concepts are easier to express visually than verbally. "Thinking with Video" allows the model to reason by generating intermediate video frames (e.g., a diagram of a geometric proof, a chart of a statistical argument), and then use the visual reasoning to inform the final answer.
The technical details: the model is trained to interleave text and video in the reasoning chain. The text provides the "narrative" of the reasoning, and the video provides the "visual evidence." The video frames are generated by a video diffusion model (Sora-2), conditioned on the text reasoning so far. The result is a "multimodal reasoning trace" that is more expressive than text-only.
The benchmark: on the MATH benchmark, "Thinking with Video" Sora-2 hits 92% — a 4-5 point improvement over text-only reasoning. On a "geometric reasoning" benchmark, the improvement is even larger (15-20 points), because geometry is naturally visual.
The bigger takeaway: "unified multimodal reasoning" is the next frontier. The "text-only reasoning" assumption is breaking, and the "video + text reasoning" approach is a significant step forward. For the industry, this means the next generation of reasoning models will be multimodal by default, and the "text reasoning" vs "video reasoning" debate will be replaced with "text + video reasoning."