Researchers built a world-model planner that decides, transition by transition, whether thinking harder is worth the computation.
A new paper describes DeepJEPA, a joint-embedding predictive world model used by AI planning systems to imagine possible futures before acting. Instead of running the same fixed number of recurrent refinement steps on every imagined transition, DeepJEPA learns to add extra computation only where it changes the outcome - mostly at the moment of contact or during sustained object interaction. Across five visual-control test settings, it matched or beat the best fixed-depth planners while averaging just 1.00 to 1.26 extra updates per transition, barely more than doing no extra work at all. The researchers also found that this improvement did not come from better overall predictions of object states, but from well-placed corrections at decision-critical moments.
Most world-model scaling efforts have leaned on brute-force approaches: more rollouts, longer horizons, deeper search trees. DeepJEPA suggests a cheaper lever - spend extra compute only where a decision is actually in play, not uniformly across every imagined step. That is a meaningful efficiency argument if it holds up outside the lab settings tested here.
It is a single-paper result on visual-control benchmarks, not a shipped product, so treat the efficiency numbers as promising rather than proven at scale.