Demystifying Video Reasoning

• Reasoning primarily follows a Chain-of-Steps (CoS) process, exploring candidate solutions early and progressively refining them into a final decision [65] [69].With bidirectional attention, the model reasons across frames simultaneously at each denoising step.Source: §1 · p. 2, Figure 1 · p. 2
• Denoising-step perturbations impair performance more than frame-wise perturbations, and conclusions primarily solidify during middle diffusion steps.Source: §1 · p. 2
• Models exhibit working memory for persistent references, self-correct errors during CoS, and localize targets before later reasoning and manipulation.Source: §1 · p. 2
• Within a diffusion step, early Transformer layers support dense perception, critical middle layers perform most reasoning, and final layers consolidate the next video state.Source: §1 · p. 2
• Training-Free Ensemble merges latents from three identical models with different random seeds to preserve diverse candidate paths and encourage convergence to correct solutions.Source: §1 · pp. 2–3

2.1 Reasoning in Language and Multimodal Models

• Research progresses from LLM reasoning and vision-language models to unified architectures supporting reasoning for generation, visual chain-of-thought, and embodied scenarios.Prior work remains focused on discrete text and static images, limiting access to spatiotemporally consistent priors.Source: §2.1 · p. 3

2.2 Video Generation Models

• Recent video generators are studied as zero-shot spatiotemporal reasoners, but the mechanisms underlying their reasoning remain unexplored.Earlier approaches instead use external reasoning modules or symbolic LLM reasoning to guide video synthesis.Source: §2.2 · p. 3

2.3 Similarities to Biological Brains

• Studies describe biological and human deliberation as internal simulation or replay of candidate trajectories before action, an analogy for diffusion-model planning.The cited examples include hippocampal replay in rats and proposed human prospective planning.Source: §2.3 · p. 3

3 Chain-of-Steps: Reasoning along Diffusion Steps

• The authors characterize reasoning as Chain-of-Steps rather than Chain-of-Frames, with denoising steps exploring, evaluating, and converging on solutions.They analyze multiple base models and VBVR-tuned reasoning-enhanced variants; visualizations primarily use VBVR-Wan2.2.Source: §3 · pp. 4–5

3.1 Diffusion Steps as the Primary Axis of Reasoning

• Decoded estimates of the clean latent reveal intermediate hypotheses, showing plausible paths emerging in parallel before suboptimal trajectories are suppressed.The estimated clean latent is the model’s instantaneous prediction of the final denoised sample, not a latent encountered during inference.Source: §3.1 · p. 5

3.1.1 Multi-Path Exploration.

• In multi-path exploration, models search parallel solution branches and progressively discard alternatives, as in maze routes and object-placement trajectories.The robot’s lower route becomes dominant, while four candidate plant trajectories collapse toward placement on the shelf’s first layer.Source: §3.1.1 · p. 5

3.2 Noise Perturbation and Information Flow

• Against the baseline without noise, Noise at Step lowers the final score from 0.685 to below 0.3, while Noise at Frame produces a much smaller performance drop.Early-step perturbations propagate throughout the trajectory and fundamentally alter the final reasoning outcome; step-wise sensitivity peaks around steps 20–30, when the reasoning trajectory is nearly finalized.Source: §3.2 · pp. 7–8

3.1.2 Superposition-based Exploration.

• In superposition-based exploration, models temporarily represent mutually exclusive states as overlapping hypotheses that resolve as denoising proceeds.Examples include overlapping circle sizes for a size-pattern continuation and a blurred superposition of multiple object orientations.Source: §3.1.2 · p. 6

3.1.3 Discussion of Reasoning Patterns.

• Among 200 inspected VBVR-Bench and real-world examples, 72% show multi-path or superposition-based exploration, with candidates more pronounced in reasoning-enhanced VBVR-Wan2.2 than general video-generation baselines.The candidates are semantically coherent and task-consistent, distinguishing them from generic blur or ghosting artifacts; simpler tasks typically converge directly.Source: §3.1.3 · p. 6

4.1 Working Memory

• Diffusion steps preserve persistent anchors, enabling the model to retain object states through occlusion and maintain consistency across a video.For example, early steps retain the small teddy bear’s state while a larger bear temporarily blocks it.Source: §4.1 · p. 8

4.2 Self-correction and Enhancement

• The model can revise incorrect hypotheses and refine incomplete answers, with reversals occurring globally across all frames within a single diffusion step.In one example, an ambiguous trajectory converges from four candidate points to one correct point.Source: §4.2 · p. 8

4.3 Perception before Action

• Early diffusion steps ground the scene by identifying target entities, while later steps introduce motion planning and coordinate object interactions.The examples show this progression for a car and a door.Source: §4.3 · p. 8, Figure 5 · p. 9

5.1 Layer-wise Token-Level Visualization

• Token-level energy maps visualize layer activations by reshaping hidden states into video-frame grids and measuring each token’s channel-wise L2 norm.The analysis extracts hidden states from transformer blocks during each diffusion step and visualizes selected layers across sequential frames.Source: §5.1 · pp. 10–11
• Within each diffusion step, activations shift from global structures and background in Layers 0–9 toward prompt-relevant objects and motion or interaction reasoning from around Layer 9 onward.This recurrent progression culminates in object-centric focus, as illustrated by a graph example whose activations ultimately concentrate on the starting node.Source: Figure 6 · p. 10, §5.1 · p. 11

5.2 Layer-wise Latent Swapping Experiment

• Swapping a middle layer (around 15 to 35) reversed object-grounding predictions across ten additional categories and five seeds, indicating semantically decisive information in middle-to-late layers.The latent-swapping experiment used object grounding at the first diffusion step with blank backgrounds and two object configurations; only one layer’s representation was replaced at a time.Source: Figure 6 · p. 10, §5.2 · p. 11

6 Training-Free Ensemble

• TFE is a proof-of-concept motivated by early-step multi-path reasoning and reasoning-active features in mid-layers.It uses latent-level ensembling of the same model, based on the hypothesis that the reasoning manifold shares a probabilistic bias toward correct outcomes.Source: §6 · pp. 11–12
• TFE averages representations from three different-seed forward passes across mid-layers 20 to 29 at the first diffusion step, filtering seed-specific noise.The spatial-temporal latent aggregation is intended to bias the model toward a more stable and logically consistent latent state.Source: §6 · p. 12
• VBVR-Wan2.2 gains 4.3% on VBVR-Bench, 35.6% on V-Reason Bench, 16.9% on MME-CoF, and 6.9% on RISE-Video with TFE.Across LTX2.3, Wan2.1, and Wan2.2 on four benchmarks, TFE consistently improves performance across nearly all model–benchmark pairs; Tab. 1 also reports improvements across VBVR-Bench subcategories.Source: Table 2 · p. 12, §6 · p. 12

7 Conclusion

• The training-free reasoning-path ensemble improves performance over a strong baseline.The conclusion presents this method as a validation of the paper’s findings.Source: §7 · p. 12
• Reasoning primarily unfolds along diffusion steps as Chain-of-Steps (CoS), challenging the hypothesized Chain-of-Frames (CoF) mechanism.This conclusion is based on qualitative analysis and targeted perturbation experiments.Source: §7 · p. 12
• The study identifies working memory retention, self-correction, and perception before action as emergent reasoning behaviors.Source: §7 · p. 12
• Middle layers of Diffusion Transformers are important for reasoning.Source: §7 · p. 12

A.1 Experiment on ensemble size

• Ensemble performance improves with additional seeds up to a best range of 3–4, then drops as excessive averaging can introduce conflicting reasoning paths.The finding indicates a trade-off between diversity and consistency in latent-space aggregation.Source: §A.1 · p. 17
• The method uses three seeds as a computationally efficient default setting.Source: §A.1 · p. 17

A.2 Experiment on Aggregation window

• 0.716 overall is the best VBVR-Bench score, achieved by aggregating layers 20–29 at diffusion step s = 0.The experiment links this reasoning-active window to more stable, informative latent representations across seeds.Source: Table 4 · p. 17, §A.2 · p. 17
• 0.688 overall for layers 0–9 is only a marginal improvement over the 0.685 baseline, with limited in-domain and out-of-domain gains.The authors suggest early layers mainly encode low-level perceptual features rather than reasoning-ready semantic structures.Source: §A.2 · p. 17
• 0.690 overall for layers 0–39 is a modest improvement, but gains are inconsistent across categories.The authors attribute this to noise from early perceptual layers and late layers already specialized for generation.Source: §A.2 · p. 17
• The comparison holds the ensemble step fixed at s = 0 and varies only the aggregated layer window.This design isolates the effect of layer range on latent ensembling.Source: §A.2 · p. 17
• The authors interpret layers 20–29 as a transition between understanding and reasoning, where semantic concepts are integrated into reasoning trajectories.They argue that ensembling in this window filters stochastic noise across seeds while preserving semantic structures that guide correct reasoning.Source: §A.2 · p. 17

B.1 Flip Rate Analysis of Layer-wise Swapping Experiments

• The swapping experiments cover ten object-category pairs and five random seeds, and the middle-layer peak persists across both.Source: §B.1 · p. 17, Figure 7 · p. 18
• Outcome-flip rates peak in the middle layers across category pairs and random seeds, forming a robust bell-shaped pattern.The pattern was observed across ten object-category pairs, each evaluated with five random seeds.Source: §B.1 · p. 17, Figure 7 · p. 18
• Elevated middle-layer flip rates suggest these layers support semantic decision-making, where competing hypotheses remain susceptible to perturbation.The authors interpret this pattern as further evidence that semantic grounding is concentrated in the network’s middle portion.Source: §B.1 · p. 18

B.2 Full Layer-wise Analysis

• The 40-block, frame-wise activation visualizations span all tasks in Sec. 5.1 and show a transition from early global structures to middle-layer foreground attention and reasoning-related motion and interaction features.Source: §B.2 · p. 18

B.2.1 High sparsity in token activations.

• Across most layers, only a small subset of spatial tokens has substantial activation, indicating sparse token-space computation concentrated in salient visual patches.This sparsity is particularly pronounced in middle layers, where large regions remain near-zero while a few tokens stay strongly active.Source: §B.2.1 · p. 18

B.2.2 High concentration of token activations in the middle layers.

• Regular grid-like activation patterns become visible in intermediate layers and may reflect patch tokenization and spatial organization.The patterns appear checkerboard- or lattice-like; the passage presents their connection to patch boundaries as a possibility.Source: §B.2.2 · p. 18

C The Impact of the Number of Frames

• Frames serve as a latent spatiotemporal workspace, or “scratchpad,” for storing essential visual information throughout diffusion.Source: C The Impact of the Number of Frames · p. 18
• A single-frame setup following ChronoEdit [69] performs substantially worse than all multi-frame settings, indicating that multiple frames support spatiotemporal coherence and video reasoning.The setup anchors input and output images at separate time steps and drops intermediate frames after a few steps.Source: C The Impact of the Number of Frames · p. 18, Table 5 · p. 21
• Reducing VBVR-Wan2.2 from ∼100 frames to around 17 causes a relatively minor performance drop, while further reductions lead to noticeable degradation.The authors interpret this as evidence that a minimum level of temporal continuity is needed to accommodate key events for correct inference.Source: Table 5 · p. 21, C The Impact of the Number of Frames · p. 21

D Performance on 4-Step Distilled Model

• The distilled-model study combines reasoning-enhancing VBVR-Wan2.2 and Phased DMD [12] speed-enhancing LoRAs, each scaled at 0.5.The evaluation uses a distilled 4-step Wan2.2-I2V-14B model.Source: D Performance on 4-Step Distilled Model · p. 21
• VBVR-Bench performance drops from 0.685 to 0.605 when the first step reduces noise too aggressively, collapsing latent exploration.Although distillation cuts denoising steps from 50 to 4, reasoning steps do not compress proportionally; preserving sufficient initial latent evolution is crucial.Source: D Performance on 4-Step Distilled Model · p. 21

E Qualitative Examples on VBVR-LTX2.3 and VBVR-Wan2.1

• VBVR-LTX2.3 and VBVR-Wan2.1 examples show Chain-of-Steps dynamics, with intermediate hypotheses refined and final conclusions stabilizing in later denoising stages.The examples also display Working Memory, Self-correction and Enhancement behaviors.Source: E Qualitative Examples on VBVR-LTX2.3… · p. 23, Figure 10 · pp. 24–25, Figure 11 · pp. 26–27

F More Qualitative Examples on VBVR-Wan2.2

• Additional VBVR-Wan2.2 examples show Multi-Path Exploration, as candidate structures are explored before converging to a coherent generation.These examples correspond to the phenomenon in Sec. 3.1.1.Source: F More Qualitative Examples on… · p. 23, Figure 12 · pp. 28–29
• VBVR-Wan2.2 examples show Superposition-based Exploration, in which multiple hypotheses may coexist in intermediate representations.These examples correspond to the phenomenon in Sec. 3.1.2.Source: F More Qualitative Examples on… · p. 23, Figure 13 · p. 30
• Further VBVR-Wan2.2 examples illustrate Working Memory preserving key information such as a trajectory.The examples correspond to the phenomenon in Sec. 4.1.Source: F More Qualitative Examples on… · p. 23, Figure 14 · p. 31
• Further VBVR-Wan2.2 examples show Self-correction and Enhancement, with early imperfect structures gradually refined through denoising.The examples correspond to the phenomenon in Sec. 4.2.Source: F More Qualitative Examples on… · p. 23, Figure 15 · p. 32, Figure 16 · p. 33