Language Models Learn Universal Representations of Numbers and Here's Why You Should Care

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Štefánik, Michal, Mickus, Timothee, Kadlčík, Marek, Højer, Bertram, Spiegel, Michal, Vázquez, Raúl, Sinha, Aman, Kuchař, Josef, Mondorf, Philipp, Stenetorp, Pontus
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866914496775192576
author Štefánik, Michal
Mickus, Timothee
Kadlčík, Marek
Højer, Bertram
Spiegel, Michal
Vázquez, Raúl
Sinha, Aman
Kuchař, Josef
Mondorf, Philipp
Stenetorp, Pontus
author_facet Štefánik, Michal
Mickus, Timothee
Kadlčík, Marek
Højer, Bertram
Spiegel, Michal
Vázquez, Raúl
Sinha, Aman
Kuchař, Josef
Mondorf, Philipp
Stenetorp, Pontus
contents Prior work has shown that large language models (LLMs) often converge to accurate input embedding for numbers, based on sinusoidal representations. In this work, we quantify that these representations are in fact strikingly systematic, to the point of being almost perfectly universal: different LLM families develop equivalent sinusoidal structures, and number representations are broadly interchangeable in a large swathe of experimental setups. We show that properly factoring in this characteristic is crucial when it comes to assessing how accurately LLMs encode numeric and other ordinal information, and that mechanistically enhancing this sinusoidality can also lead to reductions of LLMs' arithmetic errors.
format Preprint
id arxiv_https___arxiv_org_abs_2510_26285
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Language Models Learn Universal Representations of Numbers and Here's Why You Should Care
Štefánik, Michal
Mickus, Timothee
Kadlčík, Marek
Højer, Bertram
Spiegel, Michal
Vázquez, Raúl
Sinha, Aman
Kuchař, Josef
Mondorf, Philipp
Stenetorp, Pontus
Computation and Language
Artificial Intelligence
Machine Learning
Neural and Evolutionary Computing
Prior work has shown that large language models (LLMs) often converge to accurate input embedding for numbers, based on sinusoidal representations. In this work, we quantify that these representations are in fact strikingly systematic, to the point of being almost perfectly universal: different LLM families develop equivalent sinusoidal structures, and number representations are broadly interchangeable in a large swathe of experimental setups. We show that properly factoring in this characteristic is crucial when it comes to assessing how accurately LLMs encode numeric and other ordinal information, and that mechanistically enhancing this sinusoidality can also lead to reductions of LLMs' arithmetic errors.
title Language Models Learn Universal Representations of Numbers and Here's Why You Should Care
topic Computation and Language
Artificial Intelligence
Machine Learning
Neural and Evolutionary Computing
url https://arxiv.org/abs/2510.26285