Correcting a learned physical invariant improves world-model rollouts
A DreamerV3 model learned an energy-like constraint from pendulum pixels, then violated it during imagined rollouts.
Bao tests frozen conservative and damped pendulum video models with a label-free invariant search. The search finds the same conserved scalar across three conservative models, but not in matched damped ones. When autonomous rollouts drift away from that scalar’s initial level set, projecting the latent state back reduces error in all three conservative cases. Matched random constraints usually make rollouts worse. ArXiv · AI/CL/LG's note
Bao tests frozen conservative and damped pendulum video models with a label-free invariant search. The search finds the same conserved scalar across three conservative models, but not in matched damped ones. When autonomous rollouts drift away from that scalar’s initial level set, projecting the latent state back reduces error in all three conservative cases. Matched random constraints usually make rollouts worse. ArXiv · AI/CL/LG's note
score 4