Hierarchical Denoising For Multi-Step Visual Reasoning

2026-07-16

AI Summary

Existing video generation models struggle to simultaneously maintain logical consistency and enable low-latency streaming output in multi-step visual reasoning. To address this, this work proposes the HDR framework, which introduces hierarchical latent variables into causal video generation for the first time. The approach organizes latent representations in a tree structure: coarse-grained layers preserve uncertainty to support global planning, while fine-grained layers progressively refine visual states. A sparse hierarchical attention pattern (SHAP) is designed to substantially reduce temporal computational overhead. Evaluated on the first hierarchical multi-step video reasoning benchmark, HDR improves success rates from 34.22% to 60.29% (+76.2%) across six out-of-distribution tasks, achieves 54.2× faster inference than bidirectional diffusion (0.70 seconds per latent), reaches 82.9% of full-data performance using only 2% of training data, and demonstrates robust world modeling and physical interaction capabilities in real-world robotic experiments.

Abstract

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/.

Problem

• multi-step visual reasoning
• video diffusion models
• logical consistency
• low-latency streaming
• hierarchical reasoning

Innovation

• Hierarchical Denoising
• Multi-Step Visual Reasoning
• Causal Video Generation
• Sparse Hierarchical Attention
• Diffusion Models