Reading the Plan Before the Picture: How Video Models Reason Through Mazes

When we talk about video diffusion models, we usually describe what they produce: high fidelity, temporally coherent clips that capture motion and, increasingly, something resembling physical intuition. What we talk about much less is how they arrive at those outputs. A recent paper from Princeton, "Video Models Reason Early: Exploiting Plan Commitment for Maze Solving" by Kaleb Newman, Tyler Zhu, and Olga Russakovsky, takes a careful look at the internal dynamics of these models and arrives at a conclusion that reframes how we should think about reasoning in pixel space. The short version, which I summarized in a tweet, is that these models commit to a plan early and spend the rest of generation polishing pixels. The longer version is more interesting, and it has direct consequences for how we allocate inference compute.

The setup: why mazes

The first thing worth appreciating about this work is the choice of testbed. Open ended video generation is notoriously hard to study scientifically because the model is free to fabricate arbitrary detail. There is no ground truth to verify against, and failures of reasoning are entangled with failures of rendering. If a generated robot fails to pick up an object, was the plan wrong, or did the model simply render the grasp poorly?

Maze solving cleanly separates these concerns. Mazes have a long pedigree as a planning testbed, going back to Tolman's cognitive map theory and running through model based reinforcement learning to modern deep RL benchmarks. They demand sequential, constraint satisfying action planning. Crucially, conditioning the video model on a fixed input frame holds the environment constant, so a wrong trajectory is unambiguously a reasoning failure rather than a rendering artifact. Ground truth solutions come for free via breadth first search, which enables fully automatic verification, and difficulty can be tuned along independent axes: grid size, path length, and obstacle density.

This is exactly the kind of controlled instrument you want when probing an emergent capability that nobody designed the model to have in the first place.

Finding one: early plan commitment

The central empirical observation is what the authors call early plan commitment. By decoding the model's intermediate predictions of the clean signal (the standard x-hat-zero estimate) at various denoising steps, they can visualize what trajectory the model "intends" long before the final video is legible.

The result is striking. Within the first few denoising steps, often before the image is even recognizable as a maze, the high level motion plan is already fixed. Later denoising steps refine visual fidelity but rarely alter the path. In their illustration with Wan2.2-14B navigating an elf to a gift while avoiding lakes, the trajectory in the green box (early steps) is essentially the same trajectory you see in the final output, while the blue box (later steps) just sharpens texture and detail.

This connects to a well established body of work on diffusion dynamics. It is by now standard knowledge that the reverse diffusion process is coarse to fine: global semantic structure crystallizes in the earliest steps while later steps handle low level detail. Cross attention map studies and internal activation analyses have shown that spatial layout and semantic identity are fixed early, and some methods exploit this by training stage specialized denoisers per noise level. This behavior has even been formalized theoretically as sharp phase transitions with bounded critical windows.

What the Princeton paper adds is the demonstration that the abstract reasoning content, not just the visual layout, lives in that early window. The "plan" is part of the coarse structure. That is a meaningful conceptual jump. It means reasoning in these models is not an iterative search that refines a candidate solution over many steps. It is closer to a single committed guess, made under heavy noise, that the rest of the process simply renders faithfully.

Finding two: path length is the bottleneck

The second finding concerns what actually makes a maze hard for these models. Intuitively you might guess that a cluttered maze with many obstacles would be the difficult case. The data says otherwise. Path length, not obstacle density, is the dominant predictor of difficulty, and there is a sharp failure cliff at roughly 12 steps of trajectory.

This is a clean, almost suspiciously clean, result, and it tells us something about the nature of the computation. If obstacle density were the limiting factor, the bottleneck would be local constraint satisfaction: avoiding walls and lakes at each step. The fact that path length dominates instead suggests the limit is on the horizon of the committed plan. The model can hold and render a coherent trajectory only up to a certain length before the single shot plan degrades. Beyond that horizon, the early commitment that usually serves the model so well becomes a liability, because it has bet everything on a plan that is too long to get right in one pass.

ChEaP: turning the analysis into method

The practical payoff is a method the authors call Chaining with Early Planning (ChEaP), built from two components that follow directly from the two findings.

• Early Planning Beam Search (EPBS). Standard best-of-N sampling fully denoises every seed and then keeps the best one. If the plan is already visible after a handful of steps, this is enormously wasteful: you spend most of your compute polishing trajectories that were doomed from the start. EPBS instead partially denoises a large pool of seeds, scores each early plan with a lightweight verifier, and reserves full denoising for only the most promising candidates. This is a direct exploitation of finding one.

• Chaining. Because there is a hard horizon at 12 steps, long mazes simply cannot be solved in a single generation. The fix is to decompose them. The model solves a short sub-problem, then reconditions on the last frame of a successful trace to continue from there, stitching short successful segments into a long solution. This is a direct exploitation of finding two.

The reported numbers are substantial. ChEaP matches best-of-N accuracy in roughly 0.3 times the diffusion steps, delivers up to 2.5 times accuracy gains on hard tasks, and lifts long horizon maze accuracy from 7 percent to 67 percent. These results hold across two strong models, Wan2.2-14B and HunyuanVideo-1.5, and across the Frozen Lake and VR-Bench datasets over more than 480 mazes. The consistency across architectures matters; it suggests early commitment is a property of the diffusion sampling regime rather than a quirk of one model.

My take

What I find most compelling here is the reframing of what "reasoning" means in a video model. The popular framing of chain-of-frames reasoning invites an analogy to chain-of-thought in language models, where intermediate steps are a scratchpad for iterative search. This paper undercuts that analogy in an important way. The denoising trajectory is not the reasoning trace. The reasoning happens almost instantaneously, in the first few steps, and the rest is rendering. If anything, the better analogy is to a fast, intuitive, System 1 style guess followed by a slow articulation of that guess into pixels.

That has a corollary the authors gesture at but that I think deserves emphasis: chaining is the closest thing these models have to genuine sequential, deliberative reasoning. Each chained generation is a fresh early commitment conditioned on the previous outcome. In other words, the only way to get multi step search out of a single shot reasoner is to externalize the search loop, exactly as ChEaP does. This is structurally similar to how we wrap language models in agentic loops to get behavior that does not emerge from a single forward pass.

The honest limitations are worth stating. Mazes are a clean but narrow domain, and it is unclear how cleanly early commitment generalizes to open ended tasks where there is no verifier and no notion of a discrete plan length. The 12-step cliff is also likely model and resolution dependent, so the specific number should not be over interpreted. And the lightweight verifier that scores early plans is doing real work; in domains without a cheap BFS oracle, building such a verifier is itself the hard problem.

Closing thoughts

The deeper message of this paper is that current video models reason more than we give them credit for, and that we have been eliciting that reasoning badly by spending compute in the wrong place. Best-of-N over fully denoised seeds is the diffusion equivalent of reading every essay to the end before deciding which student understood the prompt. Reading the thesis statement is enough.

The open questions I would most like answered: Does early commitment persist as models scale, or do larger models begin to revise plans mid-denoising? Can we train models to extend their single shot horizon past the cliff, or is the cliff a fundamental property of committing under noise? And can the early plan readout itself become a controllable interface, letting us inspect and edit a model's intended trajectory before it ever renders a frame? That last possibility, treating the noisy intermediate as a planning surface, may be the most useful door this work opens.