Thinking in Layers: When AI Learns to "Think Before Acting"
*Translated and adapted from a paper-explainer post on zhichai.net (original by Xiaokai). The paper: "Hierarchical Denoising For Multi-Step Visual Reasoning", arXiv:2607.15278, by researchers from Peking University and UC Berkeley.*
The Problem: Fast vs. Slow Visual Reasoning
Current AI video models face a dilemma on multi-step visual reasoning tasks — planning a path through a maze, solving Tower of Hanoi, or playing Sokoban. They sit between two extremes:
- Streaming autoregressive diffusion generates frames one at a time and cannot "look back." It is fast, but achieves only 34.22% success on complex reasoning tasks.
- Bidirectional diffusion sees the whole sequence and refines every frame densely. It reasons better but runs 54.2x slower, making it impractical for real-time use.
- Tree-structured hierarchy of latents. Coarse denoising layers keep multiple parallel, uncertain hypotheses alive for global planning (like a CEO weighing strategy options), while fine denoising layers progressively commit them to concrete visual states (like departments executing plans).
- SHAP (Sparse Hierarchical Attention Pattern). Instead of attending to all timesteps at every layer, each hierarchy level attends selectively to the most relevant ones — keeping computation tractable while preserving multi-step reasoning.
- Success rate: 34.22% → 60.29% (a 76.2% relative improvement over the streaming baseline)
- Average progress: 76.00 → 89.56 — even failures advance further along a coherent reasoning trajectory
- Latency: 0.70 s/latent, 54.2x faster than bidirectional diffusion
- Data efficiency: with only 2% of training data, HDR retains 82.9% of its full-data performance, versus 52.0% for bidirectional diffusion — suggesting the hierarchy helps the model learn problem *structure* rather than memorizing training data
- Qian, Z., Chi, X., Mak, C. W., et al. (2026). *Hierarchical Denoising For Multi-Step Visual Reasoning*. arXiv:2607.15278. https://arxiv.org/abs/2607.15278
- Project page: https://hierarchical-diffusion-reasoning.github.io/
The paper's core idea echoes the proverb "think thrice before acting" — but the thinking happens inside a carefully designed "layer cake" of denoising steps.
How HDR Works
HDR (Hierarchical Denoising for Visual Reasoning) encodes a human-like coarse-to-fine planning process into the video diffusion model:
This maintains streaming inference at 0.70 seconds per latent, versus nearly 38 seconds for bidirectional diffusion.
Benchmark and Results
The team built a Level-Stratified Multi-Step Video Reasoning Benchmark with six tasks: maze navigation, Tower of Hanoi, one-line drawing (Eulerian paths), sliding puzzle, Sokoban, and water pouring.
Key findings:
Toward Real Robots
The authors also deploy HDR on physical robots performing manipulation tasks — pushing objects, avoiding obstacles, planning paths — showing that hierarchical latent planning transfers to physical interaction and world modeling, not just abstract puzzles.
Takeaway
The post frames HDR as a step from "pattern matching" toward structured reasoning: coarse layers act like a working-memory backlog, fine layers decompose it into executable steps, and sparse hierarchical attention selectively focuses on what matters now. The broader implication: AI systems may eventually simulate, plan, and even "rehearse" actions internally before executing them.