SAGE: Subgoal-Conditioned Action Generation for Latent World Model Planning
SAGE improves planning for long-horizon tasks using latent world models by structuring action search via latent subgoal decomposition. It introduces a variable-length generator to predict reachable intermediate states and a conditional action generator that proposes actions based on these predicted subgoals. This method significantly boosts success rates on complex tasks like PushT and OGBench Cube.