Demystifying Video Reasoning

TL;DR

Diffusion-based video models that “solve” mazes, puzzles, and physical-reasoning prompts have been assumed to do so via a Chain-of-Frames (CoF) — reasoning unfolding sequentially across video frames. This paper challenges that and shows the mechanism is actually a Chain-of-Steps (CoS): reasoning emerges along the denoising trajectory, not the temporal axis. By decoding the predicted clean video at each diffusion step, the authors observe multiple candidate solutions held simultaneously in early steps (e.g. several maze trajectories superimposed) and progressive convergence to one final answer in later steps. They further locate functional specialization across DiT depth (early layers = perception, middle layers = reasoning, late layers = consolidation), and demonstrate a training-free improvement by ensembling latent trajectories from the same model with different random seeds.

Key claims

Reasoning in diffusion-based video models primarily emerges along the denoising steps (Chain-of-Steps), not across video frames (Chain-of-Frames) as prior work assumed [§Abstract, §1, Fig. 1].
Decoded intermediate clean-video predictions reveal explicit parallel exploration in early steps: multiple candidate paths for a robot, multiple placements for a piece, multiple end positions for a plant, simultaneous selection of two diamonds, and superimposed rotations of an object all visible before convergence [Fig. 2].
Targeted noise-perturbation probes that disrupt information flow are used to localize where reasoning happens within the denoising process, supporting the CoS hypothesis over CoF [§4 probing experiments].
Within a single diffusion step, DiTs exhibit self-evolved functional specialization across depth: early layers encode dense perceptual structure, middle layers execute reasoning, later layers consolidate latent representations [§Abstract].
Several emergent behaviors are identified as critical to model performance: working memory (persistent reference across the trajectory), self-correction and enhancement, and perception-before-action ordering [§Abstract, project page].
A training-free strategy of ensembling latent trajectories from identical models with different random seeds improves reasoning, used as a proof-of-concept that CoS dynamics can be exploited at inference time [§Abstract, §5].

Method

The paper is primarily a mechanistic / interpretability study of an existing video reasoning model (VBVR-Wan2.2 — the team’s Wan2.2 fine-tune on the VBVR suite, cf. A Very Big Video Reasoning Suite (VBVR)). The core diagnostic is to decode the predicted clean video x̂_0 at every denoising step and inspect the visible content, rather than only looking at the final output. This reveals that early steps simultaneously render all candidate solutions and later steps prune to one — a property that is structurally invisible if you only look at the final frame.

To go from observation to mechanism, the authors run two probes. First, qualitative analysis on controlled tasks (mazes with green-circle start / red-rectangle goal, domino chains with branching paths, color-piece placement, rotation alignment) where the ground-truth reasoning trace is known and the intermediate x̂_0 decodes can be compared against it. Second, noise perturbations injected at specific layers and timesteps to disrupt information flow — if reasoning lived in frames, frame-axis perturbations should dominate; instead, denoising-axis perturbations dominate. Within a single denoising step, they further probe layer-wise behavior and find an early/middle/late functional split inside the DiT. The proof-of-concept inference-time strategy then ensembles latent trajectories across random seeds, exploiting that early-step exploration is seed-dependent.

Results

The fetched content does not surface quantitative comparison tables, but the structural results are: (1) CoS is the correct mechanistic description on the controlled tasks studied; (2) decoded clean-video at intermediate steps visibly contains multiple candidate solutions, in agreement with the parallel-exploration story; (3) layer-wise probing localizes perception / reasoning / consolidation across DiT depth; (4) the random-seed-trajectory ensemble works as a training-free reasoning booster. Companion work (A Very Big Video Reasoning Suite (VBVR)) provides the leaderboard against ten video generators — the CoS analysis is on the same VBVR-Wan2.2 model family.

A near-simultaneous independent paper, Video Models Reason Early (Newman/Zhu/Russakovsky, Princeton, arXiv 2603.30043), reaches a closely related conclusion on 2D maze-solving: video diffusion models commit to a high-level motion plan within the first few denoising steps, with later denoising altering visual details but not the trajectory. The two papers together make the CoS observation hard to dismiss.

Why it’s interesting

This is the missing mechanistic argument behind DiffThinker: Towards Generative Multimodal Reasoning with Diffusion Models‘s “native parallel exploration” claim — DiffThinker showed empirically that fine-tuned image-diffusion models hold multiple candidates simultaneously in early denoising steps and consolidate late; this paper shows the same dynamic emerges naturally in pretrained video diffusion models without any reasoning fine-tuning, on the same kind of controlled tasks (mazes, puzzles). It also reframes the running conversation in Thinking with Modalities: the “diffusion-native parallel exploration” substrate isn’t a property of one fine-tuned image model — it’s the default behavior of diffusion. For Luma, the practical takeaway is that the denoising axis is where reasoning compute is best spent in a video diffusion model — chain-of-frames inference-time tricks (more frames, longer rollouts) won’t help; chain-of-steps tricks (more steps, latent ensembling, step-conditional architectures) will. The layer-wise specialization finding (early=perception, middle=reasoning, late=consolidation) also suggests that depth-targeted interventions — e.g. fine-tuning only the middle layers for reasoning tasks, or routing extra compute there — could be a cheap lever.