Language disappears

Comparing the middle-layer representations of 8 languages (EN, ZH, AR, RU, JA, KO, HI, FR) across several models, sentences on the same topic clustered closer together regardless of language.

• Sentences about photo synthesis were closer, within Hindi, to Japanese and Chinese sentences than to cooking sentences.
• The same pattern repeated across Qwen3.5-27B, MiniMax M2.5, GLM-4.7, GPT-OSS-120B, Gemma-4 31B.
• In both dense transformer and MoE architectures, language identity weakened toward the middle layers.

In a harder test, the same concept was rendered as an English explanation, a Python function, and a LaTeX formula. As a result, ½mv², 0.5 * m * v ** 2, and "half the mass times velocity squared" showed a tendency to converge on the same region in the internal space.

The conclusion is that these models operate by compressing not just language but the representational format itself into a unified geometric space. The author interprets this in connection with Sapir-Whorf and Chomsky's perspectives, and previews further verification of whether the same phenomenon holds in larger models.