Does Action Conditioning Truly Anchor Future Predictions in Causal Dynamics

World Action Models (WAMs) are widely favored for their presumed capacity to internalize causal dynamics. However, it remains unclear whether the process of predicting future frames truly equips WAMs with such causal understanding. In this work, we reveal the indispensable role of action conditioning in learning dynamics, and demonstrate that only under a multimodal training distribution can WAMs truly acquire causal dynamics, enabling them to evaluate the current state within their latent space. In vanilla WAM architectures, the action conditioning module may be absent; in such cases, the model degenerates into a dynamics-agnostic predictor whose predicted future frames and executable actions lack causal grounding. Yet, even when equipped with action conditioning, training exclusively on unimodal expert data constrains the learning of causal dynamics, as the model is restricted to predicting a single deterministic future regardless of the input action. Through extensive experiments analyzing these scenarios and variants, we demonstrate that under multimodal, suboptimal data, action conditioning enables WAMs to learn more robust causal dynamics and exhibit an implicit state-value function within their latent space.