SAGE is a prior-conditioned planner for latent world models that replaces random action-proposal initialization with structured latent subgoals. A goal-conditioned generator predicts reachable subgoals over multiple durations, which condition candidate action sequences. A frozen world model then evaluates and refines the proposals before execution. On long-horizon tasks with target offset 150, the reported success rate increases from 12.7% to 64.7% on PushT and from 26.7% to 67.3% on OGBench Cube, while maintaining strong short-horizon performance.
No heat snapshots are available in the last 24 hours.