Adaptive Latent Capacity for World Models
ALeWM is built to use only as much latent capacity as prediction and planning need, instead of fixing that width upfront.
The paper introduces a JEPA-style world model that orders predictive information into compact prefixes of a wider latent representation. Its MixSIGReg regularizer is meant to make earlier latent blocks carry more useful signal while later coordinates can fall away. The authors test it in a controlled dynamical system and goal-conditioned visual control. They report higher mean success rates than tuned fixed-width LeWM while using lower planning capacity on average. HF Daily Papers' note
The paper introduces a JEPA-style world model that orders predictive information into compact prefixes of a wider latent representation. Its MixSIGReg regularizer is meant to make earlier latent blocks carry more useful signal while later coordinates can fall away. The authors test it in a controlled dynamical system and goal-conditioned visual control. They report higher mean success rates than tuned fixed-width LeWM while using lower planning capacity on average. HF Daily Papers' note
score 4