LeWorldModel: LeCun-style JEPA trains on one GPU, plans up to 48x faster

ylecun · x · 2026-11-10

Yann LeCun has long argued that pixel-level generation wastes compute on details instead of learning how the world works; his favored JEPA architecture predicts futures in compressed representation space but has been fragile due to representation collapse.

The LeWorldModel (LeWM) paper — authored by LeCun with researchers from Mila, NYU, Samsung SAIL and Brown — introduces the first JEPA that trains stably end-to-end from pixels:

• Only two loss terms: next-embedding prediction (MSE) plus SIGReg, a regularizer forcing Gaussian-distributed latent embeddings; tunable loss hyperparameters cut from six to zero
• No stop-gradient, EMA, or pre-trained encoders
• Just 15M parameters, trainable on a single GPU in a few hours
• Plans up to 48x faster than foundation-model-based world models while staying competitive across diverse 2D/3D control tasks
• Latent space encodes physical structure, and surprise evaluation reliably detects physically implausible events

Code is available.

Original post →

More from Research

• Two GP models turn failed battery experiments into signal, additive holds 90.3% after 50k cycles — bravo_abad · 2026-11-10
• LeCun clarifies JEPA: the real split is joint embedding vs generative, not the predictor — jbhuang0604 · 2026-11-10
• Stanford AI100 2026 report ships with a section on recursive self-improvement — conitzer · 2026-11-10
• Data science professor Lenka Zdeborová may put her full MLE course on YouTube — burny_tech · 2026-11-10
• Manav Robotics unveils IMBench, a 35-task benchmark for physical common sense, at CoRL 2026 — k7agar · 2026-11-10
• RIDArena: A New Benchmark Setting for Evaluating Rogue Internal AI Deployments — jammastergirish · 2026-11-10

Research channel →