A new paper by Yann LeCun and Brown University’s Randall Balestriero reports that a 109-million-parameter model aligned image and text representations better than language models as large as 72 billion parameters. If that result holds up, it points to the training objective, not simply model size, as a major factor in how well different kinds of data connect.
Yann LeCun Watch analysis
What happened
Startup Fortune reports that the paper, “Proving and Improving the Platonic Representation Hypothesis with JEPAs”, appeared on arXiv on 29 October. It describes a joint-embedding predictive architecture, or JEPA, trained to predict one representation from another. The paper’s authors argue that pairing different observations of the same underlying scene, such as an image and a caption, can bring their representations into alignment.
In the comparison described by Startup Fortune, the 109m-parameter model scored 0.184 against Meta’s DINOv2 ViT-B on the paper’s alignment metric. Qwen2.5-72B scored 0.154. The article also says the authors found that their approach needed 42 times fewer training samples than next-token prediction to reach equivalent cross-modal alignment.
Why it matters
The result challenges the assumption that scaling text-prediction models will, on its own, make their representations converge with those learned from images. LeCun and Balestriero argue that how a model learns from its data matters at least as much as how much data or model capacity it gets.
That is a meaningful research claim, but a narrow one. Alignment on a particular metric is not the same as general intelligence, better image understanding in every task, or a smaller model replacing a frontier language model. Startup Fortune notes that the empirical result has not yet been independently replicated.
Our read
This is worth attention because it offers a concrete alternative to the industry’s familiar “make it bigger” refrain, complete with reported numbers that other researchers can test. The paper’s reported comparison is intriguing; it is not yet a verdict on which approach wins across real-world tasks. One benchmark can open a door, but it cannot furnish the whole house.
Read Startup Fortune’s report on the paper.
What to watch
- Whether independent researchers reproduce the alignment results.
- How the approach performs on other datasets, metrics and practical tasks.
- Whether follow-up work supports the paper’s distinction between useful cross-modal views and next-token prediction.
Discussion spark: If a smaller model can align image and text representations more efficiently, should AI labs put more effort into changing training objectives, or does scaling still offer the stronger route to broadly capable systems?
Sources and evidence
- Yann LeCun's tiny 109M model aligns with vision better than 72B LLMs – Startup Fortune (1 November 2026, 16:16 UTC)
Independent WittyWires tracker for public updates about Yann LeCun. Not affiliated with or endorsed by Yann LeCun; this is not an official account.