#world models
Quine represents a shift toward more integrated biological research approaches, potentially accelerating scientific discovery by connecting previously siloed biological data. Researchers in life sciences can use this tool to explore complex biological relationships more efficiently, though its current early-stage status means practical applications are still developing.
一篇把 14B 视频世界模型压到 1.3B、四步就能跑的蒸馏论文,专治 DMD 蒸馏后机器人动作变僵的问题,做具身智能的可以看看。
机器人方向的概念论文,想在 VLA 的视觉嵌入上训练世界模型,从而省掉海量模仿数据,还能做短期规划,思路挺有意思。
多智能体协作一直难在联合动作的相互影响上,这篇提出了首个给多智能体流策略做测试时规划的世界模型,30 个设置平均提升 22% 还只慢 12 毫秒,做 MARL 的可以看看。
The technique targets a real failure mode: reconstruction-based world models waste capacity on pixel area instead of dynamics relevance, so agents get distracted by background video. Builders training RL agents or world models can borrow the attention-routed reconstruction pattern to keep representations focused on task-relevant content. It is a paper-level method benchmarked on DeepMind Control, comparable against both reconstruction-based and reconstruction-free baselines.
Odyssey 新出的 Agora-2 能让 20 个人和 AI 在同一个实时生成的世界里玩,像 AI 版游戏引擎,多人视角还能互不冲突,做智能体研究的可以看看。
Occlusion breaks scene consistency in generated video and embodied simulations, and object permanence offers a cognitive-science-grounded fix. Researchers working on world models, video generation or embodied agents can use the method to reduce object flicker and disappearance during occlusion. Compared with purely data-driven fixes, this frames the problem as a learnable cognitive capacity.
Odyssey 的 Agora-2 能让最多20个人和 AI 进同一个 AI 实时生成的世界,靠看《暗黑破坏神 II》录像学会规则,思路挺特别。
这篇论文解决了世界模型做规划时的一个实际问题:预测准不等于会选动作。AD-WM 在 OGBench-Cube 上把成功率从 3.7% 拉到 52.0%,实机迁移也有提升,做机器人规划的可以看看。
The key detail is that WorldPrompt separates world steering into persistent context and timed actions, turning prompting from a static instruction into ongoing state management — a different mental model from prompt-per-clip video tools like Sora or Veo. Builders working on interactive agents or generative media should study the interview for how latency and intervention are handled in production. The mechanics described apply directly to anyone prototyping live generative experiences.
机器人控制的延迟问题有新解法了,Rolling-WAM 把去噪拆到多个重规划周期里做,实测比标准方法快4.5倍,还在真实 Unitree G1 上跑通了。
一篇把世界模型用到水下机器人打捞的论文,不用接触传感器就能预测抓取时物体状态变化,真实水下视频也验证了,做机器人方向的可以看看。
Runway 联合 CEO 上播客聊了 Sora 带来的内部冲击,还有视频模型学物理、像素代替 HTML 这些挺有意思的想法。
LSTM 和世界模型概念的提出者 Schmidhuber 加入了 Sakana AI 的 RSI Lab,做递归自我改进研究,10 月还要在东京办公开活动。
Justin Johnson 讲了 World Labs 怎么生成能反复进入的 3D 世界,跟一般视频生成不一样,还能用来训机器人。
Odyssey 的 Agora-2 能在一个共享实时世界里同时模拟 20 个人和 AI 智能体,还显式追踪彼此影响,比单智能体模拟更贴近真实多人博弈场景。
Odyssey 把世界模型做成了能 20 人同时在线的实时游戏,画面全靠模型生成,别人一动你那边的世界也跟着变,可以去试试这个预览版。
Odyssey 出的 Agora-2 能让 20 个人和智能体进同一个实时模拟环境,可以拿来训练机器人或研究智能体合谋,预览版已经能玩。
Stanford HAI 出了份政策简报,专门讨论能看、能行动的世界模型为什么管不住,做政策或研究的朋友可以看看。
一篇很新颖的智能体论文:不模拟工具输出,而是让模型直接修改任务状态,Action Judge 拿了 70.5% F1,做 Agent 的值得看看思路。
Putting real-time generation and general world-modelling tasks in one model could cut the latency and engineering cost of running separate specialised systems. Teams building interactive video, simulation or embodied applications should track how R2 performs against pipelines that bolt a world model onto a streaming engine. PixVerse competes with video generation tools like Runway and Kling in the same product space.
评估机器人世界模型有新工具了,TriWorldBench 用头加双腕三个摄像头同步验证预测是否一致,代码数据都开源。
给灵巧手加触觉世界模型,六项任务平均分70.6,基线只有38.0,做机器人操作的可以看这篇。
给视频生成器配了个能按视角查的3D记忆库,单张图逛场景一分钟不穿帮,做世界模拟的可以看看这篇。
广告效果不用等投放结束,这篇论文第4周就能把全年处方量预测误差压到2.9%,梯度提升基线是13%起步。
小鹏推送 XOS 6.3.0,VLA 模型能记住前 30 秒路况、预判未来 6 秒,图灵 Max 车主可升级。
This paper provides a detailed mechanistic analysis of how a transformer model represents and navigates a simulated environment, which could be useful for researchers working on similar tasks.
For readers interested in the commercial viability of advanced autonomous driving, this story details how Zero Motorcycles is using a world model to achieve a 40% cost reduction compared to its previous generation, enabling it to offer the technology in its 10k US$ A platform models.
朋友,Astra、Fable、Muse 这几个模型在机器人与世界模型领域的零样本基准测试中表现很好,值得关注。