Morning Edition · Monday, September 7, 2026Published at 2:22 AM EDT · New York

Two New Preprints Push World Models Toward Physical Quantities Rather Than Pixel Reconstruction

One proposes spectral targets to structure the latent space of JEPA-style predictors, the other releases a video dataset of falling coffee grounds paired with per-frame scale readings as ground truth for weight estimation.

Latent world models predict and plan using a compressed representation rather than raw pixels, which is what makes them computationally practical for control tasks. The open problem is that nothing in the training objective forces that late…