On the Identifiability of Controlled World Models
The paper pins identifiability on action coverage, not just representation learning.
It gives a theory for when action-conditioned JEPA-style world models recover the latent state and controlled dynamics under Gaussian latent states and state-dependent Gaussian behavior policies. The authors name two required conditions: spectral separation for the predictable signal, and non-degenerate conditional action variation for transition recovery. When both hold, global minimizers identify the state and transition up to an orthogonal transformation. The paper also shows that weakly excited action directions can make counterfactual errors much larger than on-policy errors. ArXiv · AI/CL/LG's note
It gives a theory for when action-conditioned JEPA-style world models recover the latent state and controlled dynamics under Gaussian latent states and state-dependent Gaussian behavior policies. The authors name two required conditions: spectral separation for the predictable signal, and non-degenerate conditional action variation for transition recovery. When both hold, global minimizers identify the state and transition up to an orthogonal transformation. The paper also shows that weakly excited action directions can make counterfactual errors much larger than on-policy errors. ArXiv · AI/CL/LG's note
score 5