arXiv:2607.15278 (posted July 16) gives a fairly hardcore solution for video diffusion models doing "multi-step visual reasoning": HDR (Hierarchical Denoising for Visual Reasoning). The core observation is direct — streaming autoregressive diffusion is fast but can't do long-horizon planning, bidirectional diffusion can plan but each full frame reconstruction is too costly, both get stuck on logical consistency. HDR's trick is to layer the video latents into a tree hierarchy: coarse-grained layers first hold several hypotheses for global planning, fine-grained layers then concretize these hypotheses step by step into specific visual states, with SHAP (Sparse Hierarchical Attention Pattern) compressing temporal attention cost in between. The numbers are tough: average success rate on 6 OOD tasks (maze, Hanoi, one-stroke drawing, sliding puzzle, Sokoban, water-pouring) goes from the baseline's 34.22 to 60.29 (relative +76.2%), average progress from 76.00 to 89.56; latency steady at 0.70 s/latent, 54.2× faster than bidirectional diffusion; with only 2% of the training data, 82.9% of the full-data performance is preserved, while bidirectional diffusion retains only 52.0%. The authors also moved the model to a real robot for embodied experiments, pushing it toward physical interaction and world modeling. The strategic significance of this event is worth discussing. Video generation models have spent the past two years competing on "can the picture be more real", but to enter the "visual foundation model" tier, the real moat is long-horizon, corrigible, plannable reasoning — HDR's "think through the plan in the latent space first, then draw it out frame by frame" paradigm will most likely become the standard move of the next wave of video-reasoning work; and "2% data preserves 82.9% performance" quietly overturns the old assumption that "visual reasoning needs massive amounts of demonstration".