Assran et al. (2025) — ARXIV-2506.09985
Key takeaways
Two stages, and the second is small on purpose. Pre-training is action-free on more than a million hours of internet video and images. Post-training adds actions — V-JEPA 2-AC — from under 62 hours of unlabelled robot video from Droid. The ratio is the argument: the physics comes from passive observation, the control comes from a thin layer on top.
It is evaluated as three different things and holds up as all three. Motion understanding (77.3 top-1 on Something-Something v2), human action anticipation (39.7 recall-at-5 on Epic-Kitchens-100, above previous task-specific models), and — after alignment to an 8B language model — video question answering (84.0 PerceptionTest, 76.9 TempCompass). One pre-trained representation, three evaluation families.
The robotics result is the one that is hard to explain away. V-JEPA 2-AC is deployed zero-shot on Franka arms in two labs, planning to image goals, with no data collected in those environments, no task-specific training and no reward. Planning in a learned latent space is what “world model” is being asked to mean here, and this is the operational form of the claim.
Standing in the anthology
The current state of a line the record otherwise does not hold. LIT-216 established the objective — predict representations, not observations — on still images. This scales it to video, at a data volume comparable to a language pretraining corpus, and shows the resulting representation supports prediction and control rather than only recognition.
Filed as extends: and not corrects:. It names no defect in I-JEPA;
it changes modality and scale, which is what the unsigned relation is for
(ADR-017).
Read against LIT-218, the two are the same question with opposite answers about method. Both ask whether a model trained on video learns enough of the world to act on it. Veo 3 gets there by generating pixels and the capability is read off the output; V-JEPA 2 gets there by refusing to generate pixels and the capability is read off a plan in latent space. The record holds both and endorses neither — that is what makes them worth holding as a pair.
Not a practice, and the gap is the usual one. No language-model training in this record uses a JEPA objective, and this paper recommends none. What would change it: a pretraining report that trains a language or multimodal model with a latent-prediction objective and reports against an autoregressive baseline at matched compute.
Audited by #121; unsourced and unread. No reading note. One of three video/self-supervised papers the audit found in this state (LIT-216, LIT-218 are the others) — a coherent cluster nobody has opened, which is a different finding from a paper read and declined.