D-JEPA Aligns Latent World Models with Decision Outcomes
September 21, 2026
D-JEPA introduces a decision-aligned latent world model that uses a permutation-equivariant operator to refine predictive geometry. This ensures latent distance accurately reflects the success probability of competing candidate actions.
HOW THIS AFFECTS YOU
●
researcherThis method addresses the decision-local prediction gap where proximity in latent space does not correlate with action success.