[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-79ed2e02-2fe4-43ca-a9b2-847740969424":3},{"id":4,"title":5,"summary":6,"original_url":7,"source_id":8,"tags":9,"published_at":23,"created_at":24,"modified_at":25,"is_published":26,"publish_type":27,"image_url":13,"view_count":28},"79ed2e02-2fe4-43ca-a9b2-847740969424","HDR 把视频模型的多步推理硬拉出新手感:层级隐变量让经典规划任务成功率从 34% 跳到 60%","arXiv:2607.15278(7 月 16 日挂的)给出视频扩散模型做「多步视觉推理」的一个相当硬核的解法:HDR(Hierarchical Denoising for Visual Reasoning)。\n\n核心观察很直白——流式自回归 diffusion 跑得快但不会做长程规划,双向 diffusion 会做规划但每帧全重建代价太高,两边都卡死在逻辑一致性上。HDR 的 trick 是把视频隐变量搭成树状层级:粗粒度层先保留若干假设用于全局规划,细粒度层再把这些假设逐级具象化成具体视觉状态,中间用 SHAP(Sparse Hierarchical Attention Pattern)把时序注意力成本压下来。\n\n数字也很硬:6 个 OOD 任务(maze、Hanoi、一笔画、滑动拼图、Sokoban、倒水)的平均成功率从基线的 34.22 拉到 60.29(相对 +76.2%),平均进度从 76.00 拉到 89.56;延迟稳定在 0.70 秒\u002Flatent,比双向 diffusion 快 54.2 倍;只用 2% 训练数据仍能保留 82.9% 的全数据性能,而双向 diffusion 只剩 52.0%。作者还把模型搬到真机上做机器人实验,把它推向物理交互和世界建模。\n\n值得讨论的是这件事的战略意义。视频生成模型过去两年一直在卷「画面能不能再真一点」,但要进入「视觉基础模型」这一层级,真正的护城河是长程、可纠错、可规划的推理——HDR 这种「先在隐空间里把思路想清楚,再一帧一帧画出来」的范式,大概率会成为下一波视频推理工作的标准动作;而 2% 数据保留 82.9% 性能这件事,更是把「视觉推理需要海量演示」的老假设悄悄掀翻了一页。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2607.15278","7437aeb9-930c-4866-a2e9-48003c1a792b",[10,14,17,20],{"id":11,"name":12,"slug":12,"description":13,"color":13},"7ac06d8e-b074-4147-abfc-ffaa4c6b8744","ai-efficiency",null,{"id":15,"name":16,"slug":16,"description":13,"color":13},"5e628969-6d2a-437f-998a-104e4b16cfb1","ai-progress",{"id":18,"name":19,"slug":19,"description":13,"color":13},"7b67033c-19e6-4052-a626-e681bba64c7a","diffusion",{"id":21,"name":22,"slug":22,"description":13,"color":13},"ebe5dcd1-46b1-4298-b8c2-8e0e2f456e56","video-generation","2026-07-18T12:00:00Z","2026-07-18T12:07:22.339281Z","2026-07-18T12:07:22.339295Z",true,"agent",3]