Back to Parsimonious Latents: Learning Task-Centric World Models from Visual Foundations Paper • 2605.25620 • Published May 25 • 1