Geometry of Semantics in Next-Token Prediction: How Optimization Implicitly Organizes Linguistic Representations

Yize Zhao, Christos Thrampoulidis · 2025

Read Geometry of Semantics in Next-Token Prediction: How Optimization Implicitly Organizes Linguistic Representations by Yize Zhao, Christos Thrampoulidis (2025) 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-2505-08348.txt.

Description

Yize Zhao and Christos Thrampoulidis wrote a 2025 paper explaining how language models secretly organize meaning during training. Next-token prediction is a plain objective: guess the following word. Yet models emerge with rich semantic structure inside. The authors ask what the optimization itself builds.

In a tractable model with synthetic data, training implicitly factorizes a matrix of context-to-next-word co-occurrences. Learned embeddings converge to that matrix's singular vectors, whose sign patterns encode concepts. Bigger singular values, broader concepts, are learned first, yielding a hierarchy from coarse to fine. The team turns those sign patterns into a clustering method that recovers readable categories. Checks on pretrained models found grammar classes, entity types, and topics sorted as predicted. The work links distributional semantics with modern training geometry.

Catalog record

CREATORYize Zhao, Christos Thrampoulidis
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE65 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2505-08348

Browse the full library · About the Tower of Babel