The authors studied LeWM, a latent world model that predicts the outcomes of candidate action sequences and selects the sequence whose terminal latent is closest to the goal latent.
Planning Along Latent Manifolds Avoids False Shortcuts to Distant Goals
A graph-based terminal cost improved goal-directed planning across four tasks without retraining the latent world model.
Big Tech
Zunzhi You · Linwei Tao · Siyu Xu · Yunke Wang · Chang Xu
The University of Sydney · Machine Learning Research, Apple
Research Digest··3 min read
You et al.
§