CAER: Causal Action Effect Reweighting for World Model Training
2026-08-31 • Artificial Intelligence
Artificial Intelligence
AI summaryⓘ
The authors explain that current models predicting how scenes change after actions often focus too much on the background and not enough on the important parts affected by the action. They propose a new training method called CAER that shifts attention to the crucial parts influenced by the action by comparing model predictions with and without the action. This method works automatically without extra data or processing and improves how well the model understands, controls, and visually represents action effects in videos. Tests show CAER produces better and more realistic predictions than older methods.
world modelsaction-conditioned video generationmean squared errorcausal effectreweightingsupervisionmodel trainingcontrollabilityphysical consistencyvideo prediction
Authors
Jianjie Fang, Xvyuan Liu, Ziyou Wang, Rongze Tang, Zhaolu Wang, Zhuohang Li, Xin Zhang, Haisheng Su, Chen Gao, Wei Wu, Xinlei Chen, Yong Li
Abstract
World models are becoming core infrastructure for embodied intelligence, with action-conditioned video generation providing controllable predictions of how scenes evolve after agent interventions. Yet existing models are commonly trained with space-time-uniform mean squared error, allowing abundant background tokens to dominate the gradient while sparse interaction dynamics remain under-optimized; such uniform fitting rewards reconstructing appearance rather than learning how actions change the world. We introduce Causal Action Effect Reweighting (CAER), a general training paradigm that redistributes supervision toward the tokens whose predicted future is causally affected by the action. CAER contrasts the model's own predictions with and without action conditioning to localize these tokens online, then normalizes the resulting effect map into a weight that preserves the total coefficient mass and changes only where it is spent. This online signal requires no external annotations or offline preprocessing, avoids additional data-processing time, and scales naturally with model and dataset size. Experiments across heterogeneous action-conditioned world-model tasks show that CAER converges to better solutions than uniform MSE training, with consistent improvements in the physical consistency, controllability, and visual quality of generated videos.