Branch-JEPA extends JEPA world models from a single latent successor to a context-weighted finite set of latent successors, retaining the complete branch set at inference. The paper studies specialization training and full-set Energy-Score training. On the official Argoverse 2 validation split with five locked seeds, full-set training improves trajectory Energy Score by 5.8–6.5% and probability-weighted trajectory distance by 9.3–10.4% over matched-K=6 assignment and transport objectives. The model also reports gains over decoder-only branching and improved verified-route existence in an OGBench graph audit.
No heat snapshots are available in the last 24 hours.