Yang Likun releases the lightest world model ever, trainable on a single GPU, planning speed increased 48x

Summary

Yann LeCun's team introduced LeWM, a 15-million-parameter, JEPA-based world model. The model is trained end-to-end and stably from raw pixels on a single GPU within hours. It uses only one active hyperparameter and the SIGReg regularizer to prevent representation collapse. In block-pushing tasks, it reports up to a 48x speedup in planning and an 18% increase in success rate compared to PLDM.

Why it matters

World models are considered the missing piece for enabling robots to plan by simulating outcomes before acting. However, previous JEPA approaches faced issues such as representation collapse, reliance on six or more hyperparameters, and dependency on large pre-trained encoders. A model that is truly trainable with minimal tuning on consumer hardware could democratize research. This would allow prototyping embodied AI to move far beyond well-funded laboratories.

Highlights

• 15 million parameters, 2 components, 2 loss terms, 1 tunable hyperparameter — training on a single GPU within hours
• SIGReg regularization eliminates representation collapse without empirical hacks; curves are smooth and reproducible across all seeds
• Pure pixel-based LeWM outperforms DINO-WM, which integrates proprioception, in block-pushing tasks and operates up to 48x faster in planning compared to previous JEPA-based methods

Critical take

The benchmark set is limited to toy 2D navigation and block-pushing tasks, where LeWM lags behind traditional methods in navigation and still requires explicit effect labels. Therefore, the narrative that it 'understands physics' relies on a single teleportation-color-change test rather than open-world generalization. The text presents a reprocessed summary using the promotional language of the arXiv paper ('most impressive', 'qualitative leap'). There is no independent reproduction. Consequently, the claims of a 48x speedup and 18% increase should be treated as single-laboratory, single-task assertions until third-party verification is available.