Semantic information links timing of language model decisions

Breakdown of Local Denoising as Semantic Speciation

Machine Learning

Summary

Generative models that create text or images make decisions in stages: first, they decide what kind of thing they're generating (like choosing a category), then they fill in the details. This paper studies when exactly these two stages happen and finds that the stage where the model needs information from far away in the input overlaps with the stage where it decides the category. The authors prove this overlap under certain natural conditions and show that as models get larger, these stages merge into one sudden change. This helps explain how models organize meaning during generation.

What this means in practice

A theory result. No direct application yet.

Authors

Guangkuo Liu, Mert Okyay, Yifan F. Zhang, Fangjun Hu, Rahul Nandkishore, Xun Gao

Abstract

The dynamics of generative models exhibit two apparently distinct temporal windows: a speciation window, in which a sample commits to a semantic class, and a nonlocality window, in which local context windows become insufficient for generation. Motivated by evidence of their near-concurrence in a variety of frontier models, we investigate their relationship through the spatial distribution of semantic information. Under a "common cause" hypothesis, we prove that the nonlocality window must lie in the speciation window. This hypothesis postulates that semantic labels explain a fraction of the correlations between distant tokens, a condition that is natural for many real datasets. We further give conditions under which both windows shrink to a single limiting time as system size grows, defining a "phase transition", and verify this behavior analytically in Gaussian mixtures. Together, these results identify conditions under which semantic information explains the concurrence of speciation and nonlocality, connecting two complementary perspectives on the emergence of semantic structure in generative modeling.