Latent Semantic Manifolds in Large Language Models

Mohamed Mabrok · 2026

Read Latent Semantic Manifolds in Large Language Models by Mohamed Mabrok (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2603-22301.txt.

Description

Mabrok reads the hidden states of language models as points on a curved semantic surface. Models compute in continuous vector spaces but must output discrete tokens. That mismatch warps meaning in ways that are poorly mapped. Mabrok equips the surface with the Fisher information metric. Tokens then act as Voronoi regions that tile the surface. He defines an expressibility gap for the distortion this tiling causes. Two theorems bound the gap for any fixed vocabulary. Tests on six transformer models confirm the predicted scaling.

Catalog record

CREATORMohamed Mabrok
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE96189
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2603.22301

Browse the full library · About the Tower of Babel