Establishing a Scale for Kullback-Leibler Divergence in Language Models Across Various Settings

Fuente: arXiv
Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Kishino, Ryo, Takase, Yusuke, Oyama, Momose, Yamagiwa, Hiroaki, Shimodaira, Hidetoshi
Format: Preprint
Veröffentlicht: 2025
Schlagworte:
Online-Zugang:
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866913046645964800
author Kishino, Ryo
Takase, Yusuke
Oyama, Momose
Yamagiwa, Hiroaki
Shimodaira, Hidetoshi
author_facet Kishino, Ryo
Takase, Yusuke
Oyama, Momose
Yamagiwa, Hiroaki
Shimodaira, Hidetoshi
contents Log-likelihood vectors define a common space for comparing language models as probability distributions, enabling unified comparisons across heterogeneous settings. We extend this framework to training checkpoints and intermediate layers, and establish a consistent scale for KL divergence across pretraining, model size, random seeds, quantization, fine-tuning, and layers. Analysis of Pythia pretraining trajectories further shows that changes in log-likelihood space, as measured by the scaling behavior of KL divergence, are much smaller than in weight space, resulting in subdiffusive learning trajectories and early stabilization of language-model behavior despite weight drift.
format Preprint
id arxiv_https___arxiv_org_abs_2505_15353
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Establishing a Scale for Kullback-Leibler Divergence in Language Models Across Various Settings
Kishino, Ryo
Takase, Yusuke
Oyama, Momose
Yamagiwa, Hiroaki
Shimodaira, Hidetoshi
Computation and Language
Log-likelihood vectors define a common space for comparing language models as probability distributions, enabling unified comparisons across heterogeneous settings. We extend this framework to training checkpoints and intermediate layers, and establish a consistent scale for KL divergence across pretraining, model size, random seeds, quantization, fine-tuning, and layers. Analysis of Pythia pretraining trajectories further shows that changes in log-likelihood space, as measured by the scaling behavior of KL divergence, are much smaller than in weight space, resulting in subdiffusive learning trajectories and early stabilization of language-model behavior despite weight drift.
title Establishing a Scale for Kullback-Leibler Divergence in Language Models Across Various Settings
topic Computation and Language
url https://arxiv.org/abs/2505.15353