JEPA world models

Open · 8 updates · since Sep 24, 2026 · updated Nov 1, 2026 AI world-models Arc · AGI

Where it stands: JEPA world models, a research line associated with Yann LeCun, have produced a steady run of papers since late September: driving and hierarchical-planning models, an 8B robot world model claiming a first scaling law, and one built to keep learning during planning. The latest claims a proof that two separately trained JEPAs, one on images and one on their descriptions, converge on essentially the same representation, with 42 times fewer training samples and a 109M-parameter text model beating LLMs up to 72B on cross-modal alignment. Watch whether outside researchers reproduce those numbers.

Latest

• 2026-11-01
  (X layer — Yann LeCun, sharing the alphaXiv paper page): the theory layer gains its capstone claim — LeCun and Randall Balestriero's paper claims a proof of the Platonic Representation Hypothesis: two JEPAs trained completely separately, one on images of a scene and one on its descriptions, never on image–text pairs, are said to converge on essentially the same underlying representation of the world under certain conditions. Claimed results: the same image–text alignment with 42× fewer training samples than next-token prediction, and a 109M-parameter text JEPA outperforming LLMs up to 72B parameters on cross-modal alignment benchmarks. So far both numbers rest only on the authors' own report. Watch: replication of the sample-efficiency and 109M-vs-72B claims, and adoption beyond the authors.

  

• 2026-10-26 (~3:20 AM ET — x.com/ylecun): the planning-representation watch gets a research datapoint — work LeCun co-authored introduces Temporal Straightening: a world model's internal representations can form curved paths making distance-to-goal misleading; training encourages straighter internal paths, yielding more stable gradient-based planning and higher success rates on goal-reaching tasks, code released. Third-party-amplified research (Simplifying AI explainer RT'd by LeCun), not a lab product — sits beside the MotionJEPA entry (Sept 24) in the V-JEPA lineage. Watch: paper venue/evals, adoption beyond the authors.
• 2026-10-16 (~3:24 PM ET, LeCun RT of NYU Center for Data Science — x.com/ylecun): AdaJEPA — NYU CDS PhD student Ying Wang with LeCun and Mengye Ren — a world model that "keeps learning after training ends," built "to adapt during planning" (Medium write-up), the JEPA lineage's continual-learning turn where it meets the Oak Lab thesis. Academic release, not a lab announcement; no benchmark claims graded yet.