Vision-language-action models (VLAs) typically map observations directly to actions, but adding a world model lets a policy anticipate future states before acting. Until now, it has been hard to tell which design choices actually matter, because existing approaches differ simultaneously in how they represent the world, how they integrate the world model, and which policy backbone they use. This paper isolates those factors by grouping prior work into four paradigms: implicit future alignment, explicit future query, a decoupled MoT future expert, and an external world predictor.

The authors instantiate each paradigm inside a unified framework and test them across in-domain, out-of-domain, long-horizon, dynamic, few-shot, and real-world settings. Their first finding is that the best world representation depends on the integration paradigm, with one exception: 2D semantic representations give robust gains regardless of paradigm. They also find that no single paradigm is universally optimal, but the decoupled MoT future expert is the strongest overall, particularly when the policy must handle out-of-domain conditions.

World modeling consistently improves policies built on standard vision-language models, while decoupled designs are more compatible with pretrained VLAs. Compared with action-only baselines, adding a world model helps across multiple settings, with the largest gains on long-horizon and dynamic tasks. The results suggest that future work should pay close attention to the choice of representation and integration style, rather than assuming one approach will work everywhere.