PaperlayerView exact revision on arXiv

Source-linked AI summary

Hierarchical Denoising For Multi-Step Visual Reasoning

Zezhong Qian, Xiaowei Chi, Chak-Wing Mak, Tianze Zhou, Ruibin Yuan, Yuhan Rui, Hengzhe Sun, Zhuoqun Wu, Yuming Li, Siyuan Qian, Sirui Han, Shanghang Zhang

arXiv:2607.15278v1cs.CV

TL;DR

Existing video models struggle to support reliable multi-step reasoning and low-latency streaming simultaneously because causal generation limits revision while bidirectional diffusion is costly. HDR organizes video latents into a tree-structured hierarchy for coarse-to-fine reasoning before streaming output, using sparse hierarchical attention to reduce temporal costs. On the multi-step video reasoning benchmark, HDR raises overall success from 34.22 to 60.29 and achieves 0.70s-per-latent latency versus 37.92s for bidirectional diffusion.

  • Problem

    Existing video models struggle to support reliable multi-step reasoning and low-latency streaming simultaneously because causal generation limits revision while bidirectional diffusion is costly.

  • Method

    HDR organizes video latents into a tree-structured hierarchy for coarse-to-fine reasoning before streaming output, using sparse hierarchical attention to reduce temporal costs.

  • Results

    On the multi-step video reasoning benchmark, HDR raises overall success from 34.22 to 60.29 and achieves 0.70s-per-latent latency versus 37.92s for bidirectional diffusion.

  • Takeaways & Limitations

    Structured multi-scale latent planning can preserve causal efficiency while recovering strong global reasoning without dense all-to-all temporal computation.

Abstract

from arXiv · show

Video models are evolving into vision foundation models, yet they still lack human-like multi-step reasoning. Streaming autoregressive diffusion models are efficient but limited in reasoning, while bidirectional diffusion enables global revision with high inference costs due to dense frame-level denoising. Both paradigms struggle to achieve logical consistency and low-latency streaming for complex reasoning tasks. We propose HDR (Hierarchical Denoising for Visual Reasoning), a unified framework that integrates hierarchical latents into causal video generation for multi-step reasoning. HDR organizes video latents into a tree-structured hierarchy, enabling coarse-to-fine reasoning before streaming output. Coarse denoising layers preserve uncertain hypotheses for global planning, while finer layers progressively refine them into concrete visual states. A sparse hierarchical attention pattern (SHAP) further reduces temporal attention costs. We introduce a level-stratified multi-step video reasoning benchmark with out-of-distribution cases, covering six tasks: maze navigation, Tower of Hanoi, one-line drawing, sliding puzzle, Sokoban, and water pouring. Compared with streaming autoregressive diffusion baselines, HDR improves success from 34.22 to 60.29 (76.2% relative gain) and increases average progress from 76.00 to 89.56, demonstrating more consistent reasoning trajectories. HDR maintains low-latency streaming at 0.70 seconds per latent, achieving 54.2 times faster inference than bidirectional diffusion. It also retains 82.9% of full-data performance with only 2% training data, compared with 52.0% for bidirectional diffusion. Real-world robot experiments further demonstrate HDR's potential for physical interaction and world modeling. Project demo: https://hierarchical-diffusion-reasoning.github.io/.

1 Introduction

HDR addresses the tension between reliable multi-step visual reasoning and low-latency streaming by performing hierarchical coarse-to-fine denoising before output. It improves benchmark reasoning performance while retaining efficient streaming through sparse hierarchical attention.

  • Introduction: HDR organizes video latents into a tree hierarchy, enabling coarse-to-fine planning before committing to frame-level streaming output.Coarse layers preserve high-level hypotheses, while finer layers refine them into concrete visual states.
  • Introduction: Streaming autoregressive diffusion commits left-to-right and limits revision, whereas bidirectional diffusion supports global revision but requires costly dense fixed-sequence denoising.This trade-off motivates a unified approach combining hierarchical reasoning with streaming generation.
  • Introduction: HDR matches denoising strength to hierarchy level and uses SHAP to restrict token communication to local and parent-level contexts.Coarse layers remain noisier to preserve multiple global plans, while finer layers receive stronger denoising to instantiate visual states.
  • Introduction: HDR improves overall success from 34.22 to 60.29 (76.2% relative gain) and average progress from 76.00 to 89.56 across six multi-step reasoning tasks.The benchmark includes maze navigation, Tower of Hanoi, one-line drawing, sliding puzzle, Sokoban, and water pouring, including out-of-distribution cases.
  • Introduction: HDR streams at 0.70s per latent, achieving 54.2× faster inference than bidirectional diffusion.The framework is designed to preserve low-latency streaming while improving logical consistency across multi-step trajectories.

2 Related Work

Prior work studies reasoning in video generation through benchmarks for structured, spatial, physical, and multi-step tasks, while streaming autoregressive diffusion targets low-latency generation through sequential temporal computation and efficiency techniques.

  • Video Model Reasoning: Video-generation reasoning benchmarks evaluate structured problem solving, spatial cognition, physical dynamics, maze navigation, and multi-step planning.The cited benchmarks include VBVR, V-ReasonBench, and VR-Bench.
  • Video Model Reasoning: Unlike video understanding with fixed inputs, video generation must construct a coherent future trajectory satisfying local and broader constraints.
  • Autoregressive Video Diffusion Models: Streaming autoregressive diffusion replaces dense full-sequence denoising with sequential temporal computation, enabling low-latency generation and efficient KV-cache reuse [11] [10] [29].Recent approaches improve this paradigm through chunk-wise rollout, queuebased denoising, AR-guided diffusion, causal attention, training–inference alignment, distillation, cache sharing, and sliding-window KV-cache acceleration.

3 Method

HDR combines low-latency streaming with revisable multi-step reasoning by organizing video latents into a coarse-to-fine hierarchy. It uses level-aware flow matching and sparse hierarchical attention to propagate multi-scale information with low temporal attention cost.

  • Sparse hierarchical attention: SHAP flattens tree tokens in coarse-to-fine autoregressive order and restricts each token to local, parent-level, and first-frame contexts, enabling sparse attention and cross-level KV-cache reuse.Each attention row has only a constant number of valid targets independent of video length, while generated key-value states are stored in a shared hierarchy cache.
  • Motivation: HDR addresses the trade-off between globally revisable bidirectional diffusion and efficient but irreversible streaming autoregression by preserving noisy coarse hypotheses for planning before finer refinement.Bidirectional diffusion supports global revision but repeatedly updates dense sequences, whereas streaming autoregression can propagate early errors without revision.
  • Hierarchical latent representation: HDR represents video latents as a tree hierarchy whose coarse tokens encode global temporal plans and whose fine tokens instantiate local visual details and final dynamics.Inference generates hierarchy levels from coarse to fine, allowing upper layers to remain noisy and revisable while lower layers refine them into concrete visual states.
  • Training objective: HDR trains each hierarchy token with a layer-wise flow-matching objective that assigns level-specific velocity fields to global planning and concrete visual refinement.The objective sums velocity regression losses across hierarchy levels and tokens, with λ_ℓ balancing their contributions.

4 Experiments

HDR is evaluated on multi-step reasoning, streaming efficiency, robustness, and robot interaction. It improves benchmark reasoning and preserves low latency, while hierarchical layers support robustness, limited-data learning, and physical-world transfer.

  • Benchmarks and metrics: The benchmark covers six tasks requiring logical consistency across multiple steps: Tower of Hanoi, maze navigation, one-line drawing, sliding puzzle, Sokoban, and water pouring.It is stratified by difficulty and designed because existing benchmarks emphasize short clips or perceptual consistency rather than complete streaming multi-step evaluation.
  • Overall comparison: HDR raises overall success from 34.22 to 60.29 and average progress from 76.00 to 89.56 versus CausalForcing, a 76.2% relative success gain.These results indicate stronger intermediate logical consistency; full-attention baselines provide dense global interaction but are less aligned with low-latency streaming.
  • Qualitative evidence: HDR resolves ambiguous decisions through hierarchical planning before fine-grained commitment, avoiding the early local failures observed for CausalForcing in Maze and One-line tasks.The qualitative comparison attributes successful completion to revisable planning rather than immediate local commitment.
  • Streaming efficiency: HDR achieves 0.70s latency per streaming step versus 0.72s for CausalForcing and 37.92s for bidirectional diffusion.Latency is measured after KV-cache initialization, showing that HDR preserves low-latency streaming while improving reasoning.
  • Robustness and ablations: With one denoising step, HDR achieves 34.72 success, while bidirectional diffusion falls to 17.78 and CausalForcing to 11.25.With 2% of training data, HDR retains 82.9% of full-data success and 97.2% of average progress, versus 52.0% and 89.5% for bidirectional diffusion.
  • Real-world robot interaction: HDR-WAM reaches a 5.47 overall score and 3.00% average success rate on RoboDojo, exceeding no-pretraining AHA-WAM and Fast-WAM baselines.Its strongest gains occur in Long-Horizon and Memory, reaching 9.85/4.75% and 6.65/4.67%, respectively, without robot-domain or embodied-interaction pretraining.

5 Conclusion · A HDR-WAM Details · A.1 Hierarchical Action Modeling

HDR combines coarse-to-fine hierarchical latents, sparse structured attention, and entropy-matched denoising to improve long-horizon video reasoning while retaining causal efficiency. HDR-WAM extends this design to embodied control by jointly denoising visual dynamics and executable action chunks from multimodal context.

  • 5 Conclusion: HDR organizes video latents into a coarse-to-fine temporal tree, applies sparse structured attention, and allocates denoising budgets using an entropy-matched principle.This design recovers much of bidirectional diffusion’s reasoning ability without full temporal attention.
  • 5 Conclusion: HDR outperforms the causal baseline and remains competitive with bidirectional diffusion across six benchmarks, while both hierarchy levels and uncertainty preservation materially affect performance.Removing coarse hierarchy levels hurts, and fully denoising every level is inferior to preserving uncertainty at upper levels.
  • 5 Conclusion: HDR shows that global reasoning in generative video modeling can use structured multi-scale latent planning instead of dense all-to-all temporal computation while preserving causal efficiency and reasoning ability.The conclusion presents this as a broader alternative for efficient global reasoning.
  • A.1 Hierarchical Action Modeling: HDR-WAM adapts HDR to embodied world-action modeling by jointly denoising visual dynamics and actions, with visual predictions conditioned on actions and actions conditioned on language, proprioception, and visual context.The video stream predicts action-conditioned future observations, while the action stream predicts executable action chunks.
  • A.1 Hierarchical Action Modeling: HDR-WAM combines episode-level global anchors with a local action-conditioned rollout and future visual landmarks; RoboDojo uses Ne = 9, Nℓ = 9, and Nf = 4.The episode view summarizes task phase and long-range progress, while the local view supplies nearby rollout context and future landmarks.
  • A.1 Hierarchical Action Modeling: Actions align only with the Nℓ−1 local visual transitions, while episode anchors and future landmarks provide context without introducing additional action targets.The action chunk is divided into Nℓ−1 groups before joint visual-action tokenization.
  • A.1 Hierarchical Action Modeling: A block attention mask separates visual denoising from action prediction: actions attend across the action chunk and clean visual conditions, while visual transitions use corresponding and prior grouped actions.Clean visual indices are excluded from the video loss, and padding is masked in the action loss.

A.2 RoboDojo Data Processing

RoboDojo samples combine shared global episode context with local action-conditioned observations and future landmarks. Three camera views are spatially assembled, then both temporal views are encoded with a shared cached video VAE.

  • Temporal Views: Each sample uses 9 uniformly sampled episode-wide RGB frames as global task-progress anchors, plus 9 consecutive local frames and 4 uniformly sampled future landmarks.The local view begins at the current frame and follows the dataset stride; landmarks span from the local-window end to the episode end.
  • Multi-Camera Processing: The top camera and two wrist-side cameras are resized, horizontally concatenated for the side views, stacked beneath the top view, and processed with standard resize, crop, and normalization.
  • Episode-Level Caching: Both temporal views share one video VAE and are cached per episode, allowing global anchors to be reused across samples without repeatedly decoding the same long video.This preserves the full episode context while reducing redundant video decoding.

B Benchmark and Eval Details … B.6 Water Pouring

The benchmark evaluates six long-horizon reasoning tasks with exact-success and partial-progress metrics, using task-specific difficulty levels, generators, and evaluators. Each task tests distinct planning, state-transition, manipulation, or trajectory-consistency requirements through legality- and progress-aware scoring.

  • B Benchmark and Eval Details: Six mixed-benchmark tasks cover spatial planning, state transitions, object manipulation, and trajectory consistency: maze navigation, Tower of Hanoi, one-line drawing, sliding puzzle, Sokoban, and water pouring.Overall scores average task-level scores equally across the six benchmarks.
  • B Benchmark and Eval Details: Evaluation reports Success / Avg. Progress pairs, where Success requires exact task completion and Avg. Progress measures partial progress tolerant of minor visual deviations.The overall score is the equal-weight average of task-level scores across the six benchmarks.
  • B.1 Maze Navigation: Maze evaluation uses solved recursive-division grids with fixed start and goal positions, requiring valid decoded routes to reach the goal; progress is route overlap with the reference path.The default set has 50 samples, spanning grid sizes N ∈{6, 7, 8, 9, 10}.
  • B.2 Tower of Hanoi: Tower of Hanoi evaluates legal disk-move sequences that place every disk on the goal rod, including 2–5-disk in-domain and out-of-distribution initializations.Progress measures overlap with the reference shortest plan.
  • B.3 One-line Drawing: One-line drawing tests self-avoiding paths across increasing board levels, requiring complete, adjacent, non-revisiting traces that remain on the target shape.Progress combines coverage, legal-transition ratio, and on-shape ratio as Aone = 0.5 Ccover + 0.3 Rlegal + 0.2 Rshape.
  • B.4 Sliding Puzzle: Sliding-puzzle evaluation decodes legal board subsequences that preserve tile inventory and reach the ordered goal, with progress combining Manhattan improvement, final-state quality, rule adherence, and observability.Levels use 3 × 3 boards for levels 2–3 and a 4 × 4 board for level 4.
  • B.5 Sokoban: Sokoban evaluates legal walking and pushing from correctly decoded states, requiring the box to reach its target without wall crossings, overlaps, or illegal actions.Progress emphasizes box movement toward the target and also incorporates final-state quality, rule adherence, and observability.
  • B.6 Water Pouring: Water pouring evaluates stationary-frame tube states under legal complete-run transfers, requiring a solved final state, conserved blocks, capacity compliance, and low unresolved-frame ratios.Progress combines rule adherence, progress, final-state quality, and observability as Awater = 0.45 Rrule + 0.35 P + 0.15 Qfinal + 0.05 Robs.

C Entropy-Matched versus Fully Denoised Hierarchies

HDR allocates denoising steps according to each hierarchy level’s effective temporal support rather than fully denoising every layer. This entropy-matched schedule preserves coarse uncertainty while giving fine layers more capacity to resolve detailed visual states, improving overall results over uniform denoising.

  • Entropy-matched schedule: Entropy matching assigns fewer denoising steps to coarse layers and more to fine layers according to their effective temporal support.The schedule reflects increasing temporal degrees of freedom at finer layers; β controls how aggressively the budget increases toward fine layers.
  • Implementation: The default HDR schedule uses effective temporal supports [5, 8, 13, 20, 32, 50] for the 21-frame setting.The final hierarchy layer is truncated by the video length, while the underlying binary hierarchy determines these effective supports.
  • Uncertainty preservation: Coarse layers preserve flexible high-level hypotheses, whereas fine layers remove more entropy to instantiate them into detailed frame-level latents.Uniform All-50 denoising prematurely collapses high-level hypotheses and limits lower layers’ ability to correct or refine them.
  • Results: 60.29 overall success and 89.56 average progress with entropy matching exceed All-50’s 58.38 success and 88.21 progress.Table 5 compares entropy-matched, fully denoised, and alternative schedules; entropy matching achieves the best overall average progress and remains competitive in overall success.

D Time Complexity Analysis · E Comparison with SOTA Closed-source Models

HDR reduces temporal attention from quadratic to linear by using a fixed-size context over a linear-size hierarchy, while retaining low-latency streaming after prefill. On reasoning benchmarks, HDR substantially outperforms off-the-shelf closed-source models under exact-completion evaluation.

  • D Time Complexity Analysis: O(N 2) per denoising step and quadratic total attention arise in bidirectional diffusion because every token attends to all N tokens at every step.This dense interaction supports global reasoning but repeatedly recomputes the full temporal attention map.
  • D Time Complexity Analysis: Streaming AR Diffusion improves streaming efficiency over bidirectional diffusion, but its attention complexity remains quadratic as the generated prefix grows.KV caching avoids recomputing previous hidden states, but the attention length still increases over time.
  • D Time Complexity Analysis: A linear-size hierarchy and fixed-size attention context provide HDR’s main computational advantage as video length increases.A binary hierarchy contains approximately 2N −1 nodes for N frame-level tokens.
  • D Time Complexity Analysis: O(KavgN) temporal attention makes HDR linear in video length, unlike the quadratic complexity of bidirectional diffusion and Streaming AR Diffusion.HDR’s hierarchy contains approximately 2N −1 tokens, and each token attends only to a fixed-size context.
  • D Time Complexity Analysis: 0.70s per latent is HDR’s subsequent streaming latency, despite a 16.19s prefill stage versus 1.48s for bidirectional diffusion and 2.44s for CausalForcing.The longer prefill is a one-time cost introduced by the hierarchical latent tree.
  • E Comparison with SOTA Closed-source Models: HDR achieves substantially higher success and average progress than off-the-shelf Wan2.6 and Veo across all reasoning tasks under the exact-completion criterion.The comparison uses the same prompts and evaluation protocol, with scores reported using success and average progress metrics.
  • E Comparison with SOTA Closed-source Models: Wan2.6 and Veo often generate visually plausible videos but frequently fail the exact multi-step constraints required by the proposed reasoning benchmarks.Because these models are closed-source, they are evaluated off the shelf without hierarchical training or benchmark fine-tuning.

F Full Table of Ablations … H.1 Qualitative Comparison with Streaming AR Diffusion

The appendix details HDR’s ablation trends and qualitative behavior, showing robustness to reduced denoising and data, benefits from deeper hierarchies, and coarse-to-fine refinement that avoids irreversible streaming errors. Qualitative cases compare HDR with CausalForcing across the six reasoning tasks and also examine residual state-consistency failures.

  • F Full Table of Ablations: HDR remains more robust than the causal baseline under reduced denoising budgets, especially in the one-step setting.Table 8 reports the corresponding task-level success and average-progress results as mean ± std.
  • F Full Table of Ablations: Adding hierarchical layers generally improves reasoning performance, whereas shallower variants remain closer to the causal baseline.The 1-layer setting corresponds to the causal baseline, and additional layers contribute to HDR’s coarse-to-fine reasoning.
  • F Full Table of Ablations: HDR degrades more gracefully than the bidirectional baseline as training data decreases, confirming robustness under limited data.The data-reduction experiment trains both models on the full, 10%, and 2% training sets.
  • G Visualization of Hierarchical Intermediate Predictions: Intermediate predictions show HDR forming high-level hypotheses at higher hierarchy levels before progressively refining them into concrete streaming states.Figure 10 visualizes clean x0 predictions at L1, L2, and L3 before noise is added, followed by the final output.
  • H Failure Case Studies: The failure-case appendix complements quantitative results with qualitative comparisons and residual analyses of state-consistency drift in maze navigation, Sokoban, and water pouring.Section H.1 compares HDR with streaming autoregressive diffusion, while Section H.2 examines representative residual failures.
  • H.1 Qualitative Comparison with Streaming AR Diffusion: CausalForcing often makes an early locally plausible mistake that becomes irreversible, causing later frames to inherit the error and end globally inconsistent.Figure 11 compares failures and successes across all six reasoning tasks, marking baseline failure points and HDR’s successful refinements.
  • H.1 Qualitative Comparison with Streaming AR Diffusion: HDR preserves revisable coarse-level hypotheses and refines them before final streaming, enabling implicit latent-space backtracking and rule-consistent solutions.This process lets HDR preserve uncertainty about multi-step structure, revise incorrect partial plans, and instantiate solutions at finer levels.

H.2 Residual Failure Modes of HDR

HDR’s residual failures are typically late-stage state-consistency drifts rather than complete planning failures, occurring when exact geometry, object identity, or terminal constraints must be preserved. Representative cases include disappearing walls in maze navigation and Sokoban, and liquid-color collapse in water pouring.

  • Water Pouring: In water pouring, grouping is nearly correct, but orange liquid collapses into red, producing a visually plausible yet semantically invalid final arrangement.This failure is shown in Figure 14.
  • Residual Error Pattern: HDR’s remaining errors are usually late-stage state-consistency drift: coarse levels encode the intended multi-step structure, while finer levels produce nearly correct states that fail on exact constraints.The failures concentrate near rollout completion, where geometry, object identity, or terminal constraints must remain consistent.
  • Maze Navigation: A plausible maze route reaches the goal corridor, but a wall disappears near rollout completion, invalidating the final state despite correct multi-step planning.This failure is shown in Figure 12.
  • Sokoban: In Sokoban, the box largely approaches its target, but a late wall-disappearance artifact makes the final scene physically inconsistent.This failure is shown in Figure 13.

I Baselines and Implementation Details

The study compares HDR with full-attention reasoning baselines, including bidirectional diffusion and VideoMAE, while implementing HDR with a hierarchical latent tree on Wan2.2-5B-TI2V. The default 125-frame configuration uses six latent hierarchy levels and an entropy-matched denoising schedule.

  • Baselines: Full-attention baselines include bidirectional diffusion and VideoMAE, enabling global temporal interactions but requiring dense sequence processing.For VideoMAE, training masks 90%–100% of tokens, while inference masks all frames except the first.
  • HDR implementation: HDR augments Wan2.2-5B-TI2V with the hierarchical latent tree described in Section 3.
  • HDR implementation: The default 125-frame HDR setting uses six latent levels sized 1, 2, 4, 8, 16, and 32 with entropy-matched denoising steps [5, 8, 13, 20, 32, 50].The schedule's derivation and ablation are provided in the appendix.
Loading 2607.15278v1…