Zezhong Qian, Xiaowei Chi, Chak-Wing Mak +9 more
Video generation models are getting good at producing frames, but ask one to reason through a puzzle like Tower of Hanoi and it tends to fall apart. There's a trade-off at play: streaming models are fast but can't plan ahead, while bidirectional diffusion can revise the whole sequence but pays a steep compute bill. HDR splits the difference by organizing video latents into a tree, coarse to fine. The rough top layers hold uncertain guesses for planning; the lower layers sharpen them into concrete states, like sketching an outline before inking the details.
On a new six-task benchmark (maze navigation, Tower of Hanoi, sliding puzzles, and more), the authors report success rising from 34.22 to 60.29 over a streaming baseline, staying fast at 0.70 seconds per latent and running much quicker than the bidirectional method. It also holds up on just 2% of the training data.
These are the paper's own numbers on its own benchmark, so check the paper for setup details and how it fares beyond these tasks.
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),…
RoGS: Adaptive Meshgrid Gaussian for Large-Scale Road Surface Mapping
arXiv (cs.AI) · July 16, 2026Parameter-efficient Prompt Tuning of Vision Foundation Model With Adaptive Focal Loss for Interpretable MCI Screening
arXiv (cs.CV) · July 16, 2026Weakly-Supervised RGB-D Salient Object Detection via SAM-driven Pseudo Annotation and State Space Interaction-based Diffusion
arXiv (cs.CV) · July 16, 2026Video = World + Event Stream