HDR Paper: Hierarchical Denoising Delivers 76% Reasoning Gains and 54x Speedup Over Bidirectional Diffusion

Peking University, HKUST, Beihang and Muka Robotics propose Hierarchical Denoising for Reasoning (HDR), a framework that organizes video latents into a coarse-to-fine tree hierarchy for streaming autoregressive video diffusion. Reports a 76.2% relative gain in multi-step reasoning success (34.22 to 60.29) at 0.70s per latent, a claimed 54.2x speedup over bidirectional diffusion. Introduces a Sparse Hierarchical Attention Pattern (SHAP) and a six-task visual reasoning benchmark spanning maze, Tower of Hanoi and Sokoban.