首页 > AI前沿 > Counterfactual World Models: JEPAs That Reason About "What If"

Counterfactual World Models: JEPAs That Reason About "What If"

arXiv机器学习 2026-11-19 01:24 15 阅读 查看原文
用户现在需要把给定的新闻正文排版成语义化的HTML片段,要遵循给出的规则,首先每个自然段落用

,要保留所有内容,重要句子可以加?不对,先看内容,首先核心的几个部分:首先第一段是讲世界模型的反事实问题的基础?不对,先拆分自然段?不对,原文的内容:首先开头讲World models...follows: 这部分,然后它的回答区分坏决策和坏运气,然后说证明了基于事实过渡训练的世界模型自己无法回答,然后讲物理系统的充分条件,然后讲动作和扰动的交互,然后得到反事实世界模型,然后讲时长h的误差界,然后是验证部分,然后是 hindsight planning的部分?不对,要合理划分,首先先处理标题级别的?不对,原文的正文是纯文本,首先要划分小节吗?哦,用户说长文用

/

划分小节。那首先,第一个大的部分可以讲问题与核心方法?不对,先捋原文的逻辑: 1. 首先是引入反事实问题:World models operate in a world they cannot fully observe. When a hidden disturbance, such as a current pushing a boat, changes the outcome of an action, a counterfactual question follows: what would have happened under a different action, subject to the same disturbance? Its answer separates bad decisions from bad luck. 这是第一部分?不对,然后接下来是We prove that world models trained on factual transitions cannot answer it on their own, whatever their data or capacity. 这是一个核心结论,要加?然后For many physical systems, the observed outcome and a simple property of the dynamics suffice: natura non facit saltus, nature makes no leaps. 这里是核心原理。然后接下来When actions and disturbances act as separate forces, their interaction is second-order in time, so the counterfactual outcome is, to first order, the observed outcome plus the difference between the world model's predictions under the alternative and factual actions. 这是推导部分,然后This yields a counterfactual world model with no counterfactual data, extra parameters, or retraining. 这也是核心结论,加。然后接下来是误差分析部分:For transitions of duration $h$, we prove an $O(h^2)$ bound on the counterfactual error relative to the disturbance variance, establish a matching minimax lower bound from unpaired factual data, and characterize failures at discontinuities, coordinate folds, and disturbances that scale the command. 这部分可以作为一个小节?然后是实验验证部分:We validate these predictions from pixels with JEPA world models on frozen DINOv2 and masked autoencoder (MAE) encoders. 这部分,然后接下来是实验结果:On four rendered simulators, our approach improves counterfactual $R^2$ from $38$--$46\%$ to $86$--$96\%$ with the same world models; on a fifth, an underwater vehicle, it reaches $90$--$94\%$ with sixteen times more factual data. Four further frozen encoders, DINOv3, SigLIP~2, CLIP and EVA-02, give $81$--$95\%$ on the same simulators. 然后是后视规划的部分:In hindsight planning, it finds the action that would have reached the goal in $35$--$69\%$ of events, against $14$--$23\%$ for an oracle planner that knows the simulator but not the outcome. 哦,对了,要划分

小节对吧?那可以分几个小节: 第一个

问题提出与核心结论

?不对,要准确,然后内容用

。不对,等下,要保留所有内容,不能增删改。那: 首先,第一个小节可以是

反事实问题的提出

,然后下面的

是:World models operate in a world they cannot fully observe. When a hidden disturbance, such as a current pushing a boat, changes the outcome of an action, a counterfactual question follows: what would have happened under a different action, subject to the same disturbance? Its answer separates bad decisions from bad luck. 然后第二个

理论推导与方法

,然后第一个

:We prove that world models trained on factual transitions cannot answer it on their own, whatever their data or capacity. For many physical systems, the observed outcome and a simple property of the dynamics suffice: natura non facit saltus, nature makes no leaps. 不对,原文的natura non facit saltus是斜体?不对,用户没说加em,就正常?不对,用户说重要句子用,其他不要改。哦对,不能增删改内容,所以不要随便加em。那继续: 然后接下来是:When actions and disturbances act as separate forces, their interaction is second-order in time, so the counterfactual outcome is, to first order, the observed outcome plus the difference between the world model's predictions under the alternative and factual actions. This yields a counterfactual world model with no counterfactual data, extra parameters, or retraining. 然后第三个

误差界与理论分析

,然后

For transitions of duration $h$, we prove an $O(h^2)$ bound on the counterfactual error relative to the disturbance variance, establish a matching minimax lower bound from unpaired factual data, and characterize failures at discontinuities, coordinate folds, and disturbances that scale the command. 然后第四个

实验验证

,然后第一个

We validate these predictions from pixels with JEPA world models on frozen DINOv2 and masked autoencoder (MAE) encoders. 然后下一个

On four rendered simulators, our approach improves counterfactual $R^2$ from $38$--$46\%$ to $86$--$96\%$ with the same world models; on a fifth, an underwater vehicle, it reaches $90$--$94\%$ with sixteen times more factual data. Four further frozen encoders, DINOv3, SigLIP~2, CLIP and EVA-02, give $81$--$95\%$ on the same simulators. 然后第五个

规划应用效果

,然后

In hindsight planning, it finds the action that would have reached the goal in $35$--$69\%$ of events, against $14$--$23\%$ for an oracle planner that knows the simulator but not the outcome. 不对,等下,有没有需要调整的?比如,有没有列表?原文没有列表,没有引用,没有代码,所以不用那些标签。然后,重要的句子比如那两个核心结论可以加,对吧?用户说重要句子可用,是可以的。那再检查有没有保留所有内容,有没有增删? 等下,再核对原文的每一句话: 1. World models operate in a world they cannot fully observe. When a hidden disturbance, such as a current pushing a boat, changes the outcome of an action, a counterfactual question follows: what would have happened under a different action, subject to the same disturbance? Its answer separates bad decisions from bad luck. 都在第一个

里。 2. We prove that world models trained on factual transitions cannot answer it on their own, whatever their data or capacity. 加,后面接For many physical systems, the observed outcome and a simple property of the dynamics suffice: natura non facit saltus, nature makes no leaps. 都在

里。 3. When actions and disturbances act as separate forces, their interaction is second-order in time, so the counterfactual outcome is, to first order, the observed outcome plus the difference between the world model's predictions under the alternative and factual actions. 然后This yields...那个句子加,都在

里。 4. For transitions of duration $h$, we prove...那段都在

里。 5. We validate these predictions...那段在

里。 6. On four rendered simulators...那段在

里。 7. In hindsight planning...那段在

里。 然后小节的

的标题是不是可以?哦,等下,用户的小节标题是我们自己加的?不对,不对,用户说“保留原文所有信息,不得增删改内容”,哦!哦对!不能加内容,那

的标题如果是原文没有的,是不是不行?哦,我刚才犯了个错!对哦,用户说不得增删改内容,那小节标题不能自己瞎编,因为那是新增的内容?那怎么办?哦,不对,用户说“长文用

/

划分小节”,那原文的长文,怎么划分?不对,原文的内容是连贯的,难道不需要加