After 600 experiments, here's what I discovered about JEPA world models
Summary
A detailed examination of JEPA world models and LeWM, including 600 experiments showing planning improvements for long-horizon tasks without retraining. The article introduces four methods—subgoal planning every 25 steps, a latent memory graph with a cost on every jump, filling memory gaps with transformed memories, and adjusting the goal cost—and discusses their impact on success rates and remaining challenges, with references to foundational work by LeCun and colleagues.