Reproduce “Hierarchical Denoising For Multi-Step Visual Reasoning” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Zezhong Qian, Xiaowei Chi, Chak-Wing Mak, Tianze Zhou, Ruibin Yuan, Yuhan Rui, Hengzhe Sun, Zhuoqun Wu and 4 more
Researchers at Peking University and The Hong Kong University of Science and Technology developed Hierarchical Denoising for Visual Reasoning (HDR), a framework that enables robust multi-step visual reasoning in video generation models while maintaining low-latency streaming by integrating hierarchical latent spaces. HDR achieved a 76.2% relative gain in success rate on a multi-step reasoning benchmark and demonstrated comparable streaming latency to prior autoregressive methods, outperforming both streaming and full-attention baselines.
Key insights
- Combining hierarchical latent spaces with differentiated denoising strengths effectively bridges the gap between the strong reasoning of bidirectional models and the efficiency of autoregressive models.
- Maintaining high-level plans in partially noisy coarse layers allows for flexible revision and consistent multi-step reasoning without requiring dense, full-sequence updates.
- A carefully designed sparse hierarchical attention pattern efficiently supports cross-level information sharing and planning within an autoregressive generation pipeline.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.