JEPA-TTT cuts latent world-model prediction error 83% under dynamics shifts
JHU · hf · 2026-10-06
Johns Hopkins researchers present JEPA-TTT, which persistently test-time-trains the latent dynamics predictor of a pretrained action-conditioned JEPA world model.
- Self-supervised updates accumulate across episodes while the visual encoder and reward head stay frozen, preserving pretrained representations and task objectives.
- Planning requires neither a goal image nor online environment reward.
- Dense replay builds prediction windows at every temporal offset, stores them in a growing buffer, and samples minibatches for predictor updates.
- Across eight dynamics shifts in four continuous-control environments, JEPA-TTT reduced autoregressive latent prediction error by 83% on average and improved planning performance by 153% over the frozen baseline after 500 test-time episodes.
More from Embodied
- $3,500 local AI PC under fire: only 24GB VRAM and priced below its own parts cost — BLUECOW009 · 2026-10-06
- Parts alone cost ~$4,500 but Ghost's Core AI PC sells for $3,499 — where's the catch? — BLUECOW009 · 2026-10-06
- Actual raw lidar image from Waymo's 6th-gen sensor suite shared online — reed · 2026-10-06
- RT-SAFE benchmark: 94.1% of 8 frontier VLMs reach goals, only 0.7% finish with zero safety events — Lianhuiq · 2026-10-06
- Vibe Robotics: Artifact Arena makes frontier models engineer competing robots — kaixhin · 2026-10-06
- PointWAM: 3D world action model beats dexterous manipulation SOTA by 11.7 points — Chunghyun Park · 2026-10-06