LeCun and Balestriero paper explains JEPA convergence, 42x sample gap
A new arXiv paper by Yann LeCun and Randall Balestriero sets out the conditions under which independently trained JEPA models form the same internal representation, and reports a large sample-efficiency edge over next-token prediction.
Yann LeCun, a New York University professor and founder of AMI Labs, and Randall Balestriero, a Brown University professor, published a paper on arXiv on October 29, 2026, titled on the Platonic representation hypothesis using JEPA, according to aitimes.com.
The paper gives a mathematical account of the conditions under which independently trained JEPA (Joint Embedding Predictive Architecture) models form the same internal representation, up to rotation. The authors report that a JEPA using the "Good Views" data construction reached internal-representation similarity with 21.6 million training samples that next-token prediction needed 900 million samples to reach, a roughly 42-fold difference. Measured in text tokens used for training, the gap was 14-fold.
A JEPA with 109 million parameters trained on about 940 million web images with linked text scored higher on internal-representation similarity to an image AI than language models of up to 72 billion parameters. On the main evaluation, the JEPA scored 0.184 versus 0.154 for Qwen2.5-72B.
The researchers built a prediction law from the share of good-view data and the number of training samples, and used it to predict the results of 85 experiments not yet run. Errors were similar to run-to-run variation from different random seeds, though the paper notes they grew at some large training scales.
Linking image and text models
In a separate experiment, independently trained image and text models were linked by a simple mathematical transformation to generate sentences describing image content, using 30,000 image-caption pairs to learn the transformation; another experiment aligned the internal representations of the two models without image-text correspondence data, the paper says. The authors acknowledge that the experiments did not fully satisfy the mathematical assumptions and that many texts in the large-scale experiments were generated by image-captioning AI, a factor that lowered representation alignment performance when excluded. The paper drew comment from the research analysis outlet Esscher and discussion on Hacker News.
Quick answers
What is the main claim of the LeCun and Balestriero paper?
The paper gives a mathematical account of when independently trained JEPA models form the same internal representation, up to rotation, and reports that a JEPA using the "Good Views" data construction reached a given representation similarity with about 42 times fewer training samples than next-token prediction.
How did the JEPA compare with Qwen2.5-72B?
A JEPA with 109 million parameters trained on about 940 million web images with linked text scored 0.184 on the main evaluation, compared with 0.154 for Qwen2.5-72B.
Where was the paper published?
The paper was published on arXiv on October 29, 2026.