arXiv:2611.24778 · 中英对照阅读
反事实世界模型:能够推理“如果……会怎样”的联合嵌入预测架构(JEPAs)
Counterfactual World Models: JEPAs That Reason About “What If”
中文速览
当行动结果会被看不见的风、洋流或设备误差影响时,普通世界模型只能预测平均结果,无法回答“同一次干扰下,换个行动会怎样”。论文先证明,仅靠普通的事实数据,这种针对单次事件的反事实本来就无法确定;但对许多短时、平滑的物理过程,可以把实际观测到的偏差加到模型对替代行动的预测上,借此估计同一干扰下的结果。理论上,这种方法在适用条件下误差随时间步长按平方缩小,而且在只有未配对事实数据时已达到最优速度;实验中,它让多个视觉世界模型的反事实预测准确率从约四成提升到八至九成,并明显改善事后找出“当时该怎么做”的成功率。它的重要性在于,让模型能把一次失败拆分成决策失误还是运气不好,并在不收集反事实数据、不重新训练模型的情况下支持更可靠的复盘与规划。
摘要
世界模型在一个无法完全观测的世界中运行。当隐含扰动(例如推动船只的水流)改变了某个动作的结果时,一个反事实问题随之而来:在受到相同扰动的情况下,如果采取不同的动作,结果会如何?答案将糟糕的决策与坏运气区分开来。我们证明,无论数据量或模型容量如何,仅凭事实转移训练的世界模型都无法独立回答这一问题。对于许多物理系统,观测到的结果和动力学的一个简单性质就已足够:自然不作跳跃,亦即自然不会突变。当动作与扰动作为彼此独立的力发挥作用时,它们的交互在时间上是二阶的,因此,反事实结果在一阶近似下等于观测结果,加上世界模型对备选动作与事实动作的预测之差。由此,我们得到一个无需反事实数据、额外参数或重新训练的反事实世界模型。对于持续时间为 $h$ 的转移,我们证明了相对于扰动方差的反事实误差 $O(h^{2})$ 上界,基于非配对事实数据建立了相匹配的极小极大下界,并刻画了模型在不连续点、坐标折叠以及扰动会缩放指令时的失效情形。我们在冻结的 DINOv2 和掩码自编码器(MAE)编码器上,使用联合嵌入预测架构(JEPA)世界模型,基于像素验证了这些预测。在四种渲染模拟器上,我们的方法在使用相同世界模型的情况下,将反事实决定系数(R²)从 $38$ – $46\%$ 提高到 $86$ – $96\%;在第五种模拟器——一辆水下航行器上,使用多出 16 倍的事实数据后,决定系数达到 $90$ – $94\%。另外四种冻结编码器 DINOv3、SigLIP 2、CLIP 和 EVA-02,在相同模拟器上也达到 $81$ – $95\%。在事后规划中,该方法在 $35$ – $69\%$ 的事件中找到了本可抵达目标的动作;相比之下,知道模拟器但不知道结果的预言机规划器仅在 $14$ – $23\%$ 的事件中做到这一点。
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. 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. 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. 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. We validate these predictions from pixels with JEPA world models on frozen DINOv2 and masked autoencoder (MAE) encoders. 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. 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.
术语表
- world model
- 世界模型
- counterfactual
- 反事实
- factual transition
- 事实转移
- hidden disturbance
- 隐含扰动
- individual counterfactual
- 个体反事实
- counterfactual world model
- 反事实世界模型
- residual reuse
- 残差复用
- residual correction
- 残差校正
- separate action and disturbance channels
- 动作与扰动分离通道
- second-order interaction
- 二阶交互
- mixed derivative
- 混合导数
- minimax lower bound
- 极小极大下界
- unpaired factual data
- 非配对事实数据
- identifiability
- 可识别性
- hindsight planning
- 事后规划
- joint-embedding predictive architecture (JEPA)
- 联合嵌入预测架构(JEPA)
- self-supervised encoder
- 自监督编码器
- DINOv2
- DINOv2
- DINOv3
- DINOv3
- masked autoencoder (MAE)
- 掩码自编码器(MAE)
- SigLIP 2
- SigLIP 2
- CLIP
- CLIP
- EVA-02
- EVA-02
- coefficient of determination (R²)
- 决定系数(R²)
- disturbance observer
- 扰动观测器
- offset-free model predictive control
- 无偏差模型预测控制
- iterative learning control
- 迭代学习控制
- structural causal model
- 结构因果模型