All or None: Identifiable Linear Properties of Next-token Predictors in Language Modeling

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Marconato, Emanuele, Lachapelle, Sébastien, Weichwald, Sebastian, Gresele, Luigi
Format: Preprint
Published: 2024
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866912275946799104
author Marconato, Emanuele
Lachapelle, Sébastien
Weichwald, Sebastian
Gresele, Luigi
author_facet Marconato, Emanuele
Lachapelle, Sébastien
Weichwald, Sebastian
Gresele, Luigi
contents We analyze identifiability as a possible explanation for the ubiquity of linear properties across language models, such as the vector difference between the representations of "easy" and "easiest" being parallel to that between "lucky" and "luckiest". For this, we ask whether finding a linear property in one model implies that any model that induces the same distribution has that property, too. To answer that, we first prove an identifiability result to characterize distribution-equivalent next-token predictors, lifting a diversity requirement of previous results. Second, based on a refinement of relational linearity [Paccanaro and Hinton, 2001; Hernandez et al., 2024], we show how many notions of linearity are amenable to our analysis. Finally, we show that under suitable conditions, these linear properties either hold in all or none distribution-equivalent next-token predictors.
format Preprint
id arxiv_https___arxiv_org_abs_2410_23501
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle All or None: Identifiable Linear Properties of Next-token Predictors in Language Modeling
Marconato, Emanuele
Lachapelle, Sébastien
Weichwald, Sebastian
Gresele, Luigi
Machine Learning
Artificial Intelligence
Computation and Language
We analyze identifiability as a possible explanation for the ubiquity of linear properties across language models, such as the vector difference between the representations of "easy" and "easiest" being parallel to that between "lucky" and "luckiest". For this, we ask whether finding a linear property in one model implies that any model that induces the same distribution has that property, too. To answer that, we first prove an identifiability result to characterize distribution-equivalent next-token predictors, lifting a diversity requirement of previous results. Second, based on a refinement of relational linearity [Paccanaro and Hinton, 2001; Hernandez et al., 2024], we show how many notions of linearity are amenable to our analysis. Finally, we show that under suitable conditions, these linear properties either hold in all or none distribution-equivalent next-token predictors.
title All or None: Identifiable Linear Properties of Next-token Predictors in Language Modeling
topic Machine Learning
Artificial Intelligence
Computation and Language
url https://arxiv.org/abs/2410.23501