Robot world models fail when action representations change

The paper's core insight isn't just that world models fail on different action formats (that's almost expected). It's that the learned latent geometry becomes nearly orthogonal rather than gracefully degrading, meaning the model has learned to answer fundamentally different questions depending on how actions are written, not just different scales of the same question.

This connects directly to the recurrent state compression work from last week on what information RNNs can safely forget. Both papers expose a common pattern: learned representations are brittle to seemingly minor engineering choices because the model has implicitly baked those choices into its abstraction layer. The action parameterization problem is the embodied AI equivalent of the predictive quotient constraint - the model's latent space optimizes for one specific input structure, and swapping that structure breaks the entire learned geometry. This also echoes the Latent Telepathy finding that decoupling perception from task logic matters; here we see the inverse problem where task logic (action representation) silently couples to perception in ways that aren't visible until deployment.

If the same researchers or follow-up work shows that fine-tuning on just 5-10% of trajectories in the new action format recovers performance to within 10% of the original, that would suggest the underlying world model is learnable but the initialization is fragile. If recovery requires full retraining, it confirms the coupling is structural rather than just a cold-start problem.