World action models (WAMs) are trained to predict the future alongside actions, but the field has been divided on whether that future must actually be generated at inference time. Explicit WAMs denoise future video into clean frames for every action chunk, while Latent WAMs discard the future entirely to speed up inference.
A new empirical study finds that this shortcut comes at a cost. Although latent WAMs match explicit ones on in-distribution tasks, they fail to retain the generalization benefits that originally motivated WAMs. The authors evaluate generalization along three axes: environmental perturbation, data efficiency, and task generalization.
Using controlled comparisons with a matched backbone, training data, and compute budget, the study reports consistent degradation across all three axes once the action expert no longer conditions on future representations. Further analysis traces the gap almost entirely to the first denoising step, suggesting the benefit comes from preparing the future rather than fully generating it.
Based on that insight, the authors propose Simple-WAM, which reduces future modeling to a single forward pass over fully noised video tokens and adapts the training-time noise schedule to this inference behavior. Across simulation and real-world tasks, Simple-WAM reportedly leads explicit WAMs in generalization while achieving efficiency comparable to latent WAMs.
The project page, paper, code, and model weights are publicly released.