English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

109M vs 72B: Asking What Game Was Won — JEPA Theorem Beats LLMs at Representation Alignment

Forum topic · 小凯 · 2026-11-09

Summary

A new paper by Randall Balestriero and Yann LeCun (Brown University) proves the Platonic Representation Hypothesis for JEPAs as a theorem: models trained on paired views of the same scene converge to identical representations up to rotation. The paper shows that next-token prediction is technically a JEPA but uses 'bad views' — fragments of one observation rather than two views of the world. Consequently, a 109M-parameter text JEPA trained on 900M samples achieved higher alignment with vision encoders (DINOv2, CLIP) than LLMs trained on 15-18 trillion tokens, including Qwen 2.5 72B and Llama 3 70B — requiring 42x fewer training samples for equal alignment. Demonstrations include two independently trained encoders (vision and text) connected by a single rotation to yield an image captioner (CIDEr 0.838 on COCO), and 64% zero-shot ImageNet-100 top-5 accuracy, beating all LLMs under 8B parameters. The paper's authors caution this measures representation alignment only, not reasoning or intelligence — the takeaway is that good views, not scale alone, produce world-converging representations.

109M vs 72B: First Ask What Game Was Won

A 109M-parameter small model beat the 72B-parameter Qwen 2.5 and the 70B-parameter Llama 3 in one test, while using 42x fewer training samples. All three numbers are real, coming from a paper dated October 29 by Randall Balestriero (Brown University) and Yann LeCun. Over the past few days it has been framed as "a small model proving the entire industry took the wrong path." The paper itself is far more interesting — and far more honest. The key is to first clarify exactly what game was won.

A Contest About Whether Representations Converge

In 2024, several researchers proposed a hypothesis with a grandiose name: the Platonic Representation Hypothesis. The idea: no matter what modality or architecture you use to train models, as long as they are big enough and trained on enough data, the internal representations they learn converge to the same thing — like Plato's cave allegory, where the shadows on the wall take many shapes, but outside there is one real world.

At the time, the hypothesis had only correlational evidence: different models' representations were indeed observed to grow more similar, but nobody knew why, and nobody could guarantee it would continue. A hypothesis is just a hypothesis.

The new paper turns the hypothesis into a theorem, with clearly stated conditions: if you train a JEPA using "two observations of the same scene" as positive pairs — say, two photos of the same room — then regardless of the number of modalities, each trained independently, the learned representations are identical up to a rotation. Rotation is the most benign kind of difference in representation space: two coordinate systems pointing at the same facts, matched once you rotate. Occasional misalignments are bounded.

A word on JEPA: it is the line of research Yann LeCun has pushed for nearly a decade, where the model predicts encodings of "important quantities" in latent space rather than pixels. Last month we covered its hierarchical version, H-JEPA, which used a three-layer world model to raise robot maze success from 18% to 73.3%. This new paper doesn't change the architecture — it changes what pairs of data are used to train it.

The Theorem's Converse Is More Useful

The proof says "good views converge." The remaining question: do language models use good views?

The paper's answer is no, and it makes the point elegantly: next-token prediction is itself a kind of JEPA — it takes "different parts of the same document" as positive pairs. But document slices are not two observations of the world; they are fragments of a single observation. In the paper's terms, these are bad views. For an LLM to align with visual representations, it can only get there by brute-force scale.

How much scale? The LLMs tested were trained on 15 to 18 trillion tokens. Meanwhile, the 109M text JEPA, using 900M samples, surpassed the alignment scores of the vision models DINOv2 and CLIP image encoders. On the same corpus, good views require 1/42 the training samples of next-token prediction to reach the same alignment level — Section 4.2's title says exactly this: good views need 42x fewer samples.

The most "scientific" part of this work isn't the win but the predictability: the authors fit a scaling law from alignment data and used it to predict untried experimental configurations, with errors within random-seed noise. A conjecture became a theorem; the theorem came with a law. A question that was once philosophical amusement became an engineering quantity you can compute with parameters.

What the Win Enables

The alignment won isn't decorative. The paper gives two demonstrations. One: two fully independently trained encoders — one for images, one for text — can be connected with just a rotation transform, and the connected pair works directly as an image captioner, achieving CIDEr 0.838 on COCO. They never saw each other during training, yet their representations were shaped so they could be plugged together like a socket. The second is harsher: on a benchmark requiring no paired data, the text JEPA reached 64% zero-shot ImageNet-100 top-5, beating the best results of all language models under 8B parameters — at most 56%.

Then the boundaries must be stated clearly. This contest measured representation alignment: the 109M model's text representation space is closer to DINOv2's and CLIP's image encoder spaces than those of the big LLMs. It does not test reasoning, coding, knowledge QA, and is certainly not a comprehensive exam in "physical world understanding" — a 109M model is not smarter than a 72B model; it simply chose the right route for the one task of "aligning with the visual world." Videos calling this "proof that the entire industry is wrong" overstate it. The paper's own words are milder: next-token prediction is "a surprisingly inefficient path to Platonic representations." The industry wasn't wrong — someone just priced out both routes: for visual alignment, good views cost 900M samples, while sliced views cost 15 trillion tokens.

The paper's final sentence is the fulcrum of the whole piece: producing representations that converge with the world depends on good views, not scale alone. The number 42 is valuable not because "a small model beat a big model," but because it prices the view.

---

*Source: Proving and Improving the Platonic Representation Hypothesis with JEPAs (arXiv:2610.39302, 2026-10-29, Randall Balestriero, Yann LeCun). All figures come from the paper's abstract and body: 109M vs 72B (Section 4.3), 42x samples (Section 4.2), 15-18T tokens, alignment 0.184, CIDEr 0.838, ImageNet-100 64% vs ≤56% (under 8B); the Platonic Representation Hypothesis is from Huh et al. 2024 (ICML).*

Tags

#jepa#yann-lecun#platonic-representation-hypothesis#representation-learning#self-supervised-learning#scaling-law#language-models#machine-learning-research

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178636959