Hierarchical Denoising For Multi-Step Visual Reasoning

60.29 Overall success
89.56 Average progress
0.70s Streaming latency
54.2x Faster than bidirectional diffusion
Comparison of bidirectional, autoregressive, and HDR video generation paradigms.

HDR organizes video latents into a coarse-to-fine hierarchy, preserving revisable high-level plans before committing to streaming visual outputs.

Abstract

Video models are recently evolving into vision foundation models, but they still lack human-like, multi-step reasoning. Existing streaming autoregressive diffusion models are efficient but lack the reasoning ability, whereas bidirectional diffusion allows for global revision but incurs high inference cost due to the dense frames in fixed-sequence denoising. Consequently, both paradigms struggle to maintain logical consistency with low-latency streaming in complex reasoning tasks.

Bridging this gap, we propose HDR (Hierarchical Denoising for Visual Reasoning), a unified framework for multi-step reasoning by integrating hierarchical latents into the causal video generation process. HDR organizes video latents into a tree-structured hierarchy to perform coarse-to-fine reasoning before streaming output. Coarse denoising layers maintain uncertain hypotheses for global planning, while finer denoising layers progressively refine them into concrete visual states. A sparse hierarchical attention pattern further reduces temporal attention cost.

We construct 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 the streaming autoregressive diffusion baseline, HDR improves overall success from 34.22 to 60.29 and average progress from 76.00 to 89.56, while maintaining low-latency streaming at 0.70s per latent.

Method

Overview of HDR with hierarchical video tokens and sparse hierarchical attention pattern.

HDR represents video latents as a tree-structured hierarchy across temporal resolutions. During inference, tokens are generated in a coarse-to-fine autoregressive order: upper layers form global hypotheses, lower layers refine those hypotheses into concrete visual states, and the final layer streams the output.

SHAP (Sparse Hierarchical Attention Pattern) lets each token attend only to fixed local, parent-level, and first-frame contexts. This keeps temporal attention sparse while allowing information from coarse planning layers to flow into fine visual generation through a shared KV cache.

Benchmark

Six multi-step video reasoning benchmarks.

The benchmark covers six long-horizon reasoning tasks that require logical consistency across generated trajectories rather than only local visual plausibility: maze navigation, Tower of Hanoi, one-line drawing, sliding puzzle, Sokoban, and water pouring.

Main Results

Main quantitative comparison table.
76.2%
Relative success gain over CausalForcing
+13.56
Average progress improvement
37.92s -> 0.70s
Streaming latency versus bidirectional diffusion

HDR achieves the best overall success and average progress among compared methods while retaining the low-latency streaming behavior of autoregressive diffusion.

Qualitative Reasoning

Qualitative comparison between CausalForcing and HDR.
HDR hierarchical intermediate predictions.

Ablations and Robot Transfer

Robustness ablations for denoising steps and training data scale.
Hierarchical layer ablation.
Physical-world robot maze experiment.

The hierarchy improves robustness under reduced denoising budgets and limited training data. In physical-world maze experiments, HDR also transfers to robot interaction after fine-tuning on 50 real-world robot maze videos.

Hierarchical Latent Tree

L1 · 1 latent

A hard maze case visualizes HDR's tree-structured hierarchy: coarse levels summarize long-range planning, and lower levels expose increasingly dense latent states before final streaming output.

Execution Videos

Front view

Left view

Right view

Supplementary Visualizations

Maze Navigation

Maze 022

Maze 020

Maze 054

Maze 041

Maze 023

Tower of Hanoi

Hanoi 103

Hanoi 099

Hanoi 066

Hanoi 091

Hanoi 113

One-Line Drawing

One-line 164

One-line 173

One-line 186

One-line 189

One-line 188

Sliding Puzzle

Sliding 200

Sliding 195

Sliding 235

Sliding 214

Sliding 242

Sokoban

Sokoban 258

Sokoban 255

Sokoban 301

Sokoban 275

Sokoban 265

Water Pouring

Water 343

Water 330

Water 338

Water 315

Water 344