Fast-WAM: Do World Action Models Need Test-time Future Imagination?

Technical Analysis Fast-WAM, a WAM architecture that retains video co-training during training but skips future prediction at test time. High-level Analogy: Imagine you're learning to play a complex strategy game, like chess. At first, you might spend a lot of time mentally simulating future moves and counter-moves (this is like how traditional World Action Models, or WAMs, 'imagine the future' at test time). This helps you learn the game's dynamics. But once you become a grandmaster, you don't explicitly simulate every single possible future move for every turn. Instead, your brain has developed a deep, intuitive understanding of the board, allowing you to quickly perceive the situation and make a strong move almost instantly. Fast-WAM is like that grandmaster: it uses its deep understanding of 'how the world works' (gained during training by learning to predict the future) to make fast, effective decisions, without needing to laboriously 'imagine' every future scenario during the actual game. Motivation of the Work Current embodied AI agents often rely on Vision-Language-Action (VLA) models that directly map observations and instructions to actions. While powerful, these models typically don't explicitly understand how actions change the physical world. To address this, World Action Models (WAMs) have emerged. WAMs are appealing because they explicitly model how visual observations might evolve under an agent's actions, which helps them capture physical dynamics and…