The JEPA stack thickens: seven papers in four days
From the AI digest for 2026-10-10
What happened: LeCun's feed is running a world-models pipeline: H-JEPA (Balestriero; hierarchical, end-to-end learned, long-horizon visual planning — Visual AntMaze 18% → 73%) and its alphaXiv page (LeCun as advisor, named); RoboJEPA (8B — "robot world models now have scaling laws," the scaling claim is new to the lineage); LeAVJEPA (sound+vision self-supervised, one shared encoder); DSReg (provable recovery of individual latent variables — a theory answer to "why does latent prediction work"); financial world models (Balestriero, NeurIPS — "financial data is RIPE for JEPAs"); and LeWAM (Oct 9): JEPAs that "jointly imagine what happens and what to do next" — 32.3× faster planning, 19 ms/plan, contact-rich manipulation success 28.6% → 89.7%.
What's genuinely new: the lineage is moving from critique-proof demos to scaling-law and theory papers — the same consolidation move transformer-era research made in 2017–2020. The claims are still academic-paper grade (vendor-adjacent; LeCun-amplified by construction). The ELLIS UnConference's world-models sessions and the Michael Jordan Libération interview circulate as community layer.
Watch: independent replications (Visual AntMaze/LeWAM numbers); whether "robot world models have scaling laws" gets a wire-grade writeup; the world-models thread now has a paper-cadence clock.
Links: LeCun's feed (RTs; no per-paper links captured) · H-JEPA arXiv 2610.06805