Schrödinger's Cat in Video Models
In 1935, Erwin Schrödinger put a cat in a box with a Geiger counter, a bit of radioactive material, a hammer, and acid. One atom might decay in an hour, or it might not. If it does, the hammer breaks the flask and the cat dies. Until you open the box, you do not know if the cat is alive or dead. According to Schrödinger, the cat is alive and dead, at the same time.
This is the point of quantum mechanics. Existence in the form of probabilities. Many science-fiction films grew out of this notion. In Coherence (2013), a comet passes over a dinner party and the house splits into many versions of itself, each going down a different trajectory. Quantum simulations on world scale have never been observed; the notion of probabilistic trajectory is only preserved to the micro-world for quantum mechanics.
90 years later, it is the best picture I have for how a video model solves a problem.
Look at the two clips. The task is simple: move the yellow ball from the blue square to the red square without hitting any X. On the right is the final video. One ball, one clear path. On the left is the same video at the second denoising step. There is no single ball. Faint copies of it appear on many paths. One is already dim near the goal. The model has not decided yet. The steps after this are the act of deciding. By the last step, the box is open and one cat walks out.
We used to think video models reason like language models: one frame after another, each based on the last. A paper I helped with this year, led by Ruisi Wang and accepted at ECCV 2026, shows this is mostly wrong. The reasoning happens across the denoising steps, not the frames. Early steps keep many options open. Later steps remove the wrong ones. This is not just a quirk of diffusion. It is what deliberation looks like in every animal that has been studied doing it.
The rat at the fork
In the 1930s, Karl Muenzinger watched rats at a maze fork. A rat would stop, turn its head toward one path, then the other. It did not move. He called this vicarious trial and error: trying paths without walking them. Edward Tolman used this idea. He imagined a sowbug that swings toward the stronger pull before it moves. A reflex acts. A mind looks both ways first.
Seventy years later, David Redish's lab recorded from rats' hippocampus as they did the same task. Place cells fire when the animal is in a specific spot. At the fork, while the rat paused and looked left and right, the place cells fired in sequences. They ran ahead of the body: down one arm, then the other. The map was testing the paths before the rat moved.
In 2013, Brad Pfeiffer and David Foster made the map visible. They recorded up to 250 place cells at once while rats searched a two-meter arena with thirty-six wells. One well was home for the day. When a rat paused, its hippocampus fired a burst lasting about a tenth of a second. Decoded, the burst showed a path from where the rat stood to the remembered home, at twenty times running speed. Then the rat ran nearly the same path. It did this even when the start and goal had never been paired before. Not a memory replay. A route drawn before it was taken.
Look at what the decoder shows during one of these events. Not a dot. A blur of probability that moves across the floor, brightest along the path, faint elsewhere. Some of that blur is the decoder's own uncertainty. I should not overread it. But watch several events in a row: different paths, some to home, some not, one after another, from the same spot. The hippocampus proposes routes. The body picks one.
That is the left panel of the figure at the top. Several ghosts of the ball, each partway down a route, none chosen. The rat proposes routes one at a time, in bursts. The model proposes them all at once and lets them fade. The shape of the act is the same: draw the path before you pay for it.
Watch the steps
If you want to see a video model think, do not watch the frames. The frames are the cat after the box is open. Watch the steps. Lay out one video's denoising path from noise to clean image. You will see routes appear, compete, and die. Just like Redish's place cells swept down one arm, then the other.
Sources
- Wang, R., Cai, Z., Pu, F., Xu, J., Yin, W., Wang, M., Ji, R., Gu, C., Li, B., Huang, Z., Deng, H., Lin, D., Liu, Z. & Yang, L. (2026). Demystifying Video Reasoning. ECCV 2026. arXiv 2603.16870 · project page
- Schrödinger, E. (1935). Die gegenwärtige Situation in der Quantenmechanik. Naturwissenschaften 23:807–812, 823–828, 844–849. English translation by J. D. Trimmer (1980), The Present Situation in Quantum Mechanics. Proc. Am. Philos. Soc. 124:323–338.
- Muenzinger, K. F. (1938). Vicarious trial and error at a point of choice: I. A general survey of its relation to learning efficiency. J. Genet. Psychol. 53:75–86.
- Tolman, E. C. (1939). Prediction of vicarious trial and error by means of the schematic sowbug. Psychol. Rev. 46:318–336.
- Pfeiffer, B. E. & Foster, D. J. (2013). Hippocampal place-cell sequences depict future paths to remembered goals. Nature 497:74–79. paper · author manuscript
- Redish, A. D. (2016). Vicarious trial and error. Nat. Rev. Neurosci. 17:147–159. · Johnson, A. & Redish, A. D. (2007). Neural ensembles in CA3 transiently encode paths forward of the animal at a decision point. J. Neurosci. 27:12176–12189.
Maze clips from Wang et al. 2026, of which the author is a co-author. Rat videos from the supplementary information of Pfeiffer & Foster 2013, reproduced with attribution for commentary.