LeCun 与 Balestriero 用数学论证柏拉图表征假说并给出更优训练路径
摘要
alphaXiv 介绍 Yann LeCun 与 Randall Balestriero 的一项工作，从数学上表明在一定条件下，两个完全独立训练的 JEPA 可以学到基本相同的世界表征，一个从同一场景的不同图像学习，另一个从该场景的不同描述学习，全程不使用图像文本配对。
摘要由模型根据原文撰写，以原文为准。
正文 · 原文
Yann LeCun and Randall Balestriero just Proved the Platonic Representation Hypothesis, and found a better way to achieve it.
The Platonic Representation Hypothesis suggests that AI models trained on different modalities, like images and text, eventually learn similar representations of the world. But this paper shows that how models are trained matters more than simply scaling them up.
So they mathematically show that, under certain conditions, two JEPAs trained completely separately can learn essentially the same underlying representation of the world. One learns from different images of the same scene, while the other learns from different descriptions of that scene, without ever being trained on image-text pairs.
Their approach achieves the same image-text alignment with 42x fewer training samples than next-token prediction. A 109M-parameter text JEPA even outperforms LLMs up to 72B parameters on cross-modal alignment benchmarks.
来源：alphaXiv · x.com
引用原文
askalphaxiv. Yann LeCun and Randall Balestriero just Proved the Platonic Representation Hypothesis, and found a better way to achieve it.[EB/OL]. (2026-11-01)[访问日期]. https://x.com/askalphaxiv/status/2116779287390896418.