Hierarchical Denoising For Multi-Step Visual Reasoning

The Problem: Video Models Can't Reason Like Humans

Video generation models today excel at mimicking patterns but fail at multi-step logical tasks. Ask a model to solve a maze or arrange blocks in sequence, and it struggles—it can't hold multiple possible futures in mind while planning ahead. Current approaches hit a wall: fast streaming models (autoregressive diffusion) generate frames one-by-one without global foresight, while slower bidirectional models that revise entire sequences are computationally expensive (requiring dense frame-by-frame refinement). Neither achieves both speed and reasoning consistency.

How HDR Works: Coarse-to-Fine Planning Then Streaming

The authors' key insight is to separate planning from output. HDR organizes video latents—compressed representations of visual information—into a tree-like hierarchy. Before streaming anything to the user, the model reasons at coarse levels first, exploring rough solutions ("maybe move the block left"), then progressively refines them toward concrete visual details.

Think of it like sketching a maze solution in light strokes, then tracing the path darker and darker. Coarse layers preserve multiple uncertain hypotheses; finer layers lock in specific actions. Once coarse reasoning settles on a plausible trajectory, only then does the model stream refined frames at low latency (0.70 seconds per latent).

A secondary innovation, sparse hierarchical attention pattern (SHAP), cuts the computational cost of attending to all past timesteps—a bottleneck in video models. Instead of dense attention, the model attends sparsely across the hierarchy, reducing temporal overhead without sacrificing reasoning quality.

Results: Dramatic Gains on Reasoning Tasks

The authors benchmarked HDR on six logical tasks (maze navigation, Tower of Hanoi, drawing, sliding puzzles, Sokoban, water pouring) with held-out test cases to measure robustness.

Compared to streaming autoregressive baselines, HDR improved:

• Success rate from 34.22% to 60.29% (76.2% relative gain)
• Average task progress from 76.00 to 89.56

These aren't marginal improvements—the model now solves most tasks end-to-end rather than getting stuck midway.

Speed vs. bidirectional diffusion: HDR is 54.2× faster (0.70 sec/latent vs. the dense alternative), eliminating the latency penalty traditionally paid for global reasoning.

Data efficiency: With only 2% of training data, HDR retains 82.9% of full-model performance—meaningfully better than bidirectional diffusion's 52.0% in the same regime.

Real-world robot experiments (moving objects in physical space) suggest the approach transfers beyond synthetic tasks, hinting at practical world modeling potential.

Why It Matters

This work addresses a genuine gap in vision foundation models: the ability to plan logically before committing to output. By decoupling coarse reasoning from fine-grained generation, HDR shows that you don't have to choose between speed and consistency. For robotics, autonomous systems, or any application requiring sequential decision-making under uncertainty, low-latency multi-step reasoning is essential. The benchmark itself—with out-of-distribution test cases—provides a clear standard for measuring reasoning generalization in video models.