AdaJEPA: An Adaptive Latent World Model
Ying Wang, Oumayma Bounou, Yann LeCun, Mengye Ren · 2026
Read AdaJEPA: An Adaptive Latent World Model by Ying Wang, Oumayma Bounou, Yann LeCun, Mengye Ren (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2606-32026.txt.
Description
Wang and three colleagues make latent world models keep learning while they are being used. Such models predict future states in a compact latent space and support planning through model predictive control. Ordinarily the model is frozen after training, so when test conditions shift, through lighting changes or altered friction, its predictions drift and the planner optimizes the wrong objective.
AdaJEPA instead adapts inside the control loop itself. After executing the first planned action chunk, the model treats the observed transition as a self-supervised signal, updates itself, and replans with the refreshed model. As few as one gradient step per replanning step substantially improves planning success across goal-reaching tasks. Gains are strongest under distribution shift and when training data is limited.
Catalog record
| CREATOR | Ying Wang, Oumayma Bounou, Yann LeCun, Mengye Ren |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 45 KB |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2606.32026