CausalWAM, a new extension of World Action Models (WAMs), learns to isolate the hidden variables that actually cause robot actions, rather than merely copying observed behaviour. By adding a sparse routing step that forces the action head of a pretrained WAM to rely on only a few latent variables, the model can identify the action‑relevant block of variables up to a permutation.

The method keeps the full world‑modeling objective of the original WAM while adding a sparsity penalty to the action head. During training, each action hidden state selects a small set of clean context indices, and only those indices contribute to the action prediction. This design preserves world‑prediction accuracy and guarantees that the action mechanism depends on a minimal, interpretable set of latent variables without extra assumptions about temporal structure.

When tested on realistic robot tasks—RoboCasa365, LIBERO, and a real‑robot stacking challenge—CausalWAM outperformed in most skill families, though it did not improve performance on every task that has the same training data and recipe but lacks the sparse router. Across all skill families and on the real robot, CausalWAM achieved higher mean success rates, demonstrating that the action‑relevant latent variables truly matter for closed‑loop control.

Further evidence shows that using only the routed tokens identified by CausalWAM retains most of the performance gains, whereas random token selection drops success sharply. For example, on four RoboCasa365 tasks, CausalWAM achieved an average success of 71.3 %, C‑only (routed tokens) 60.3 %, and random tokens 24.5 %.

A key limitation is that CausalWAM can identify the block of action‑relevant latents but cannot pin down each individual latent variable. This lack of component‑wise identifiability means that the method does not yet support precise interventions on single latent factors.