Identifying Latent Actions and Dynamics from Offline Data via Demonstrator Diversity

In three linesTheoretical paper on recovering latent actions and environment dynamics from offline trajectories without action observations. Authors leverage demonstrator diversity (each following distinct policies) to identify latent transition kernels via nonnegative matrix factorization. Identifiability proven under rank and policy diversity conditions.

Summary generated by Claude — human-verified