LeJEPA proves identifiable world models

This is a serious theoretical step for world models, but the guarantee is narrower than the hype suggests: it is mathematically clean inside a Gaussian, additive-noise setting, not a universal recipe for any environment.

• –The central result is linear identifiability up to rotation, which is exactly the kind of structure you want if a latent model is meant to support planning and compositional generalization
• –The Gaussian uniqueness claim is the sharpest part of the paper: it says the regularizer is not just a training trick, it selects the latent geometry that makes recovery possible
• –The Lean 4 formalization matters here because it reduces the usual “theory paper” ambiguity and makes the proof claims harder to hand-wave away
• –The empirical section supports the story with synthetic mixings and a pixel-based Reacher task, but it is still far from broad embodied deployment
• –The next real test is action-conditioned, partially observed worlds, where identifiability and planning often stop lining up cleanly