Yann LeCun JEPA Paper Says a 109M Model Aligns With Vision Better Than 72B Language Models

When AI researchers talk about scale, they almost always mean more parameters and more training tokens. A paper posted to arXiv on October 29 by Brown University computer scientist Randall Balestriero and Yann LeCun argues that, at least for one long-debated question, scale is the wrong lever. In "Proving and Improving the Platonic Representation Hypothesis with JEPAs," the two researchers, who both list AMI Labs as an affiliation, present a theorem stating that whether separately trained AI models converge on the same representation of the world depends on how their training pairs are built, not on how large the models are. The work is a preprint and has not yet been peer reviewed.

The experimental result attached to that theorem is what drew attention. A 109-million-parameter text model trained with the authors' approach scored .184 on the cross-modal alignment benchmark used to establish the original hypothesis, ahead of Qwen2.5-72B at .154 and Llama-3-70B at .151. The authors note that the comparison language models have roughly 660 times more parameters and were trained on 15 to 18 trillion tokens, while their encoder read 116 billion. On a shared corpus, their method also reached the alignment of next-token prediction with 42 times fewer training samples.

What the Platonic Representation Hypothesis Claims

The new paper builds on a 2024 ICML position paper by Minyoung Huh, Brian Cheung, Tongzhou Wang and Phillip Isola. The Platonic Representation Hypothesis argued that representations inside deep networks are converging. Surveying earlier work and running their own measurements, the authors showed that as vision models and language models get larger, they measure the distance between data points in increasingly similar ways, even though they are trained on different data with different objectives.

Huh and his co-authors hypothesized that this convergence is heading toward a shared statistical model of reality, which they likened to Plato's concept of an ideal reality behind appearances, and named it the "platonic representation." They discussed several selective pressures that might drive the trend and laid out limitations and counterexamples to their own analysis.

What the 2024 paper did not provide was a proof. Balestriero and LeCun write that without one, the hypothesis offers neither a guarantee that convergence will happen nor a recipe for producing it. Their measurements also show why scale alone looks like a weak recipe: on the original benchmark, Qwen2.5 models from 7 billion to 32 billion parameters score between .118 and .130, and the 72-billion-parameter version reaches only .154.

LeCun and Balestriero's October paper attempts to supply both the guarantee and the recipe.

What Good Views Are and Why Next-Token Prediction Uses Bad Ones

The paper's central idea is a distinction between "good views" and "bad views," two kinds of positive training pairs a self-supervised model can be taught to treat as equivalent.

A good-view pair is two separate observations of the same scene. For an image model, that means two photographs of one scene, or several frames from the same video. For a text model, it means two different captions written about the same image. The paper's own illustration pairs "A red sphere next to a blue cone" with "A blue cone behind a red ball": different words, same underlying scene.

A bad-view pair is two corruptions of a single observation. Two masked copies of the same sentence are bad views. So is a sentence prefix and its continuation, which is exactly the input-output relationship in next-token prediction. The problem, the authors argue, is that corruptions of one observation share that observation's incidental details, such as a caption's exact word choice or a photograph's lighting and viewpoint. A model trained on them spends part of its representation encoding those modality-specific "nuisances," which have no counterpart in any other modality. Two observations of the same scene share only the scene, so a model trained on good views spends nothing on nuisance.

The authors formalize this as a Platonic Representation Theorem. In an idealized setting, Joint-Embedding Predictive Architectures (JEPAs) trained separately on any number of modalities learn the same representation up to a single rotation whenever their positive pairs are good views, and their misalignment is bounded otherwise. A rotation, formally an orthogonal transformation, preserves every distance and angle in an embedding space, so two models that differ only by a rotation agree on which data points are neighbors and on all cosine similarities. They encode the same geometry, just in different orientations.

The second theoretical result is aimed squarely at large language models. The paper shows that next-token prediction has the form of a JEPA trained on bad views, which, the authors argue, explains why language models align with vision models only through brute-force scale.

Where JEPA Came From

LeCun proposed the Joint-Embedding Predictive Architecture in 2022 as an alternative to generative training. Instead of reconstructing raw pixels or tokens, a JEPA uses a context encoder to read part of an input and a predictor to estimate the internal representation of another part. Because the prediction happens in embedding space, the model can ignore low-level detail and is rewarded for capturing more abstract structure.

Meta's FAIR lab has built a series of models on the idea. I-JEPA, described in a January 2023 paper on which LeCun is a co-author, learned image representations by predicting the representations of target blocks from a single context block; the team trained a ViT-Huge/14 model on ImageNet with 16 A100 GPUs in under 72 hours. Meta introduced V-JEPA for video in February 2024. V-JEPA 2, published in June 2025, was pretrained on more than 1 million hours of internet video, reported state-of-the-art human action anticipation on Epic-Kitchens-100 and, as V-JEPA 2-AC, planned robot pick-and-place tasks after post-training on less than 62 hours of robot video. Meta's V-JEPA 2 repository added V-JEPA 2.1 checkpoints in March 2026, ranging from 80 million to 2 billion parameters.

Balestriero and LeCun have also pushed JEPA into language before. With Hai Huang, they released LLM-JEPA in September 2025, a JEPA-based objective for fine-tuning and pretraining large language models. In November 2025 the pair published LeJEPA, a training recipe that the new paper's experiments use. What is new in the October paper is the cross-modal question: not whether a JEPA can learn useful language features, but whether separately trained JEPAs for text and images are guaranteed to land on the same representation.

Read more: JEPA-Anything applies one world model across seven sciences

The 660x Parameter Gap and What It Actually Measures

The headline comparison uses the benchmark from the 2024 Platonic Representation Hypothesis paper. Each text model embeds 1,024 Wikipedia image captions from the WIT dataset, a fixed vision model, Meta's DINOv2, embeds the matching images, and the score measures mutual nearest-neighbor agreement: for each item, how many of its 10 closest neighbors in the text model's space are also among its 10 closest neighbors in the vision model's space. Balestriero and LeCun followed the released code of Huh and colleagues for these scores.

On that test, their 109-million-parameter text JEPA, a BERT-Base-sized encoder trained on the texts accompanying 940 million web images, reached .184. Every language model the authors evaluated scored lower, up to Qwen2.5-72B at .154 and Llama-3-70B at .151, as did the text encoder of OpenAI's CLIP at .124, even though CLIP was trained jointly with images. The paper reports that the JEPA also leads the language models when scored against other image encoders, including MAE, I-JEPA and CLIP's own image encoder, for example .230 against .210 with CLIP ViT-H, and that the result holds across two benchmarks and two alignment metrics. The authors also stress that their model never saw pixels.

Understanding what this means, and what it does not, is essential.

The score measures how similarly two models organize the same set of paired examples. It does not mean a 109-million-parameter encoder can answer questions, write code or solve math problems better than Qwen2.5-72B, and the paper makes no such claim. What the text JEPA has learned is an internal geometry that sits closer to the geometry a vision model learned independently from images.

The authors then test whether that geometric agreement is usable. In one experiment they join frozen DINOv2 and DINOv3 image encoders to a language model fine-tuned on text alone, using one linear map with orthonormal rows as the only component fitted on image-caption pairs. No neural network is trained on paired image-text data, yet the resulting captioner scored CIDEr .838 on COCO. In a second experiment, the rotation is found with no pairs at all, purely from the geometry of unpaired images and captions. On that pair-free benchmark, from Schnaus and colleagues, the text JEPA reached 64% top-5 zero-shot accuracy on ImageNet-100, against at most 56% previously reported with language models of up to 8 billion parameters.

Training Efficiency: Why Positive-Pair Design Matters More Than Data Volume

The 42x figure comes from a controlled comparison in which good-view JEPAs, bad-view JEPAs and next-token prediction were trained on the same texts with the same encoder. A good-view JEPA reached .109 on the benchmark after 21.6 million training samples; next-token prediction reached .111 only after 900 million. Because each good-view sample reads three texts, the advantage narrows to 14 times fewer text tokens, and the good-view run saw 435 times fewer unique images. At an equal 21.6 million samples, the gap is wider: .108 for the good-view JEPA against .071 for next-token prediction with the 33-million-parameter encoder.

The paper also reports a scaling law for alignment, expressed in terms of the fraction of good views and the number of samples seen. According to the authors, it predicted the results of 85 new training runs before they were trained, as accurately as rerunning an experiment with a different random seed.

From those results, the authors conclude that next-token prediction, the objective behind today's largest models, is "a remarkably inefficient route to platonic representations." If alignment between models depends mainly on how positive pairs are constructed, then additional compute spent scaling next-token prediction is an expensive way to close the gap between language and vision representations.

The paper is candid about the practical catch. Good views for text require knowing which texts describe the same scene. On Flickr and COCO, that comes from human captions grouped by image. On Recap-DataComp, the web-scale corpus behind the 109-million-parameter result, two of each image's three texts were written by an image captioner, LLaVA built on Llama-3-8B with a CLIP ViT-L/14 vision encoder, so those texts carry information distilled from image-text training. Without any captioner, good views built from alt-text and a text-only paraphrase reach .060 on the benchmark versus .107 with the recaptions. The authors write that their comparisons with language models and CLIP at scale rely on these texts, while the comparisons between view types, which use identical texts, do not.

AMI Labs and What the Paper Represents

The paper arrives with institutional weight. LeCun announced in November 2025 that he would leave Meta at the end of the year after 12 years, five as founding director of FAIR and seven as chief AI scientist, to build a startup around his Advanced Machine Intelligence research agenda. That company, Paris-based AMI Labs, raised $1.03 billion in March 2026 at a $3.5 billion pre-money valuation, with Cathay Innovation, Greycroft, Hiro Capital, HV Capital and Bezos Expeditions co-leading and Nvidia, Samsung, Temasek and Toyota Ventures among other backers.

LeCun is AMI Labs' co-founder and chairman; Alexandre LeBrun is chief executive. The company has said it is building world models that learn from reality rather than from language alone, and TechCrunch reported that it does not plan to generate revenue for now, with Nabla named as the first partner to access early models.

Balestriero joined Brown's computer science department as an assistant professor in summer 2024 after a PhD at Rice University and a postdoctoral fellowship at Meta's FAIR. His research centers on practical deep learning theory, and the October paper follows LLM-JEPA and LeJEPA in a run of JEPA papers he has co-authored with LeCun since September 2025.

Both authors list AMI Labs as an affiliation on the paper, which ties a formal result directly to the research bet the company was funded to pursue: that predicting in representation space, with the right training pairs, is a more efficient path to models that understand the world than ever-larger next-token predictors.

Read more: Meta FAIR maps scaling laws for robot world models

Caveats and What Independent Replication Would Need to Confirm

All results are self-reported in a preprint that was days old at publication of this article, and they have not yet been peer reviewed or independently reproduced.

The authors list several limitations themselves. Their theorems are population-level statements about an idealized, finite world in which every observation fully determines its underlying state, while real modalities are lossy; a caption describes only part of a photograph. Uniqueness up to rotation is proved for a simpler anti-collapse condition than the LeJEPA objective actually used in training, and the authors acknowledge that no experiment meets the theorem's hypotheses exactly. The image side mostly uses crops of one photograph as positives, which by the paper's own definition makes those image encoders, including DINOv2, bad-view JEPAs. Hyperparameters were tuned only for the good-view JEPA.

The headline number also comes with qualifications the authors spell out. A fourth training seed with a different regularization weight reached only .121; averaged across all four seeds, the model scores .168, still above every language model tested. The recaptioner behind most of the training texts sees images through CLIP ViT-L/14. And at least 43 of the 1,024 benchmark captions also appear among the training corpus's alt-texts, although the authors report that removing them changed none of the nine text models they checked by more than .002.

The reframing of next-token prediction as a bad-view JEPA is as much an interpretive argument as a technical one, and other researchers may contest whether the analogy holds for production language models trained on far more diverse objectives.

What the October 29 paper does deliver is a formal theorem where the Platonic Representation Hypothesis previously rested on an empirical trend, controlled evidence that the predicted efficiency advantage appears at practical scales, and downstream demonstrations that the alignment can be put to use. Whether it changes how large models are trained, or remains a research-track alternative to the dominant paradigm, will depend on independent replication and on what AMI Labs publishes next.