Sora-2 solves visual puzzles by drawing its reasoning in video

When we ask a model to reason, it reasons in words if the medium is text, or over a static scene if the medium is an image. The world, though, is not static: objects move, and the rules often only become visible in how those objects behave over time. The authors propose video generation as a general-purpose channel for reasoning. Text can be written directly into the frames, visual hypotheses can be drawn, then checked, then corrected. What comes out is a workspace that holds logic, geometry and language in a single process.

What they did

The authors assembled a benchmark of tasks called VideoThinkBench. It contains vision-centric tasks (assorted visual puzzles, plus mazes and ARC-AGI-2) and text-centric ones (subsets of the familiar GSM8K, MATH, MMLU/Pro, GPQA). It also folds in existing multimodal benchmarks such as MathVista and MMMU, adapted, of course, for video evaluation.

Tasks are answered through two channels. In the video, the model is asked not only to give a final answer but to show its work — draw or underline something, write out the formulas, state the answer explicitly. In the audio, it says the answer and nothing else. Scoring is done separately: from the last frame, or from a transcript of what was spoken. For visual tasks, the video brings new checkable rules: did the model draw a line connecting the two points, does that line touch the walls, did it complete the missing piece of the shape correctly and does the piece match the reference.

How the model handled purely visual tasks

On visual tasks Sora-2 often comes out ahead of the rest. On geometry it does well at finding the intersection of two rays, the center of a circle, a reflection, the path to a point and so on. The authors note that Sora-2 can drop perpendiculars. The best scoring method, by their observation, is not the final frame alone but a majority of frames: that stability strips out the random artifacts at the end of the video.

In visual puzzles, Sora-2 does well on continuing a color or shape pattern — filling in the right region, picking the right size for a figure, drawing the reflection correctly. The authors note that this looks rather unexpected for a video generator, because it takes real induction and not just template reproduction.

ARC-AGI-2 is a harder story. Strict automatic grading returns a fairly low score, but manual review turns up plenty of partially correct answers, along with cases where Sora-2 changes its solution mid-video — apparently an attempt at self-correction. On ARC-AGI-2, scoring a majority of frames instead of a single one and adding a vote across several runs lifts final accuracy noticeably.

The maze set deserves separate mention as an interesting stress test. It checks whether Sora-2 can carry its maze-solving skill across topologies. The authors note that Sora-2 traces the shortest path through a maze with some confidence when the maze sits on a square grid, but gets badly lost when the maze is laid out on a hexagonal or circular one. The skill, in other words, does not fully transfer to new scenes.

Where these abilities come from

The authors ran additional contamination tests. They generated GSM8K and MATH problems that appear in no training set but are structurally similar. Results were comparable, which lowers the odds that the system is simply recalling solutions. Another important finding concerns a hidden prompt rewriter. On the related Wan2.5 system, switching the rewriter off nearly zeroes out performance on reasoning tasks, and switching it back on restores it. That hints at a component inside that takes a short request and rewrites it into a detailed prompt for step-by-step solving. It would also explain what the worked-out video is for in Sora-2: it supplies robustness, holding the pipeline together even when the prompt has been rewritten, while the audio works well because it does not have to record every step, only the final answer. Even with a hidden rewriter inside, the system still has to execute that step-by-step plan properly, pushing complex actions in video and language into every task.

What all this is for

We are used to thinking of text as the space for reasoning, of images as the domain of visual recognition, and of video as a storytelling tool. But Sora-2 uses video as a workspace for solving problems, which gives it the following:

• Test hypotheses by acting — draw a line, erase a mask, complete a shape.
• Hold language and vision together — write the formulas and point at the diagram at the same time.
• Use time for self-correction, since several passes with a vote can get past the noise.

The approach is far from perfect: skill transfer across geometries leaves much to be desired, and part of the performance may rest on the internal prompt rewriter. It is exciting all the same. This is a single environment where world-modeling, vision and language skills can coexist not as a set of tokens but as actions in time and space.

The main takeaways: video as a single medium for reasoning works, and the system learns both to answer and to act. Sora-2 is especially strong on tasks that call for spatial or inductive reasoning, and it gains from stability across frames and from repeated runs. On the text sets, audio beats video.

Moving forward will take better tools for recording the reasoning inside the video, and training that rewards transfer across geometries. That is what makes thinking through video generation a natural next step for AI: a single reasoning environment where language, perception and action live in one shared stream of time.

💾 Code

📊 Benchmark