Saved in:
Bibliographic Details
Main Authors: Papadopoulos, Vassilis, Wenger, Jérémie, Hongler, Clément
Format: Preprint
Published: 2024
Subjects:
Online Access:https://arxiv.org/abs/2401.17505
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866909266073354240
author Papadopoulos, Vassilis
Wenger, Jérémie
Hongler, Clément
author_facet Papadopoulos, Vassilis
Wenger, Jérémie
Hongler, Clément
contents We study the probabilistic modeling performed by Autoregressive Large Language Models (LLMs) through the angle of time directionality, addressing a question first raised in (Shannon, 1951). For large enough models, we empirically find a time asymmetry in their ability to learn natural language: a difference in the average log-perplexity when trying to predict the next token versus when trying to predict the previous one. This difference is at the same time subtle and very consistent across various modalities (language, model size, training time, ...). Theoretically, this is surprising: from an information-theoretic point of view, there should be no such difference. We provide a theoretical framework to explain how such an asymmetry can appear from sparsity and computational complexity considerations, and outline a number of perspectives opened by our results.
format Preprint
id arxiv_https___arxiv_org_abs_2401_17505
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle Arrows of Time for Large Language Models
Papadopoulos, Vassilis
Wenger, Jérémie
Hongler, Clément
Machine Learning
Artificial Intelligence
Computation and Language
We study the probabilistic modeling performed by Autoregressive Large Language Models (LLMs) through the angle of time directionality, addressing a question first raised in (Shannon, 1951). For large enough models, we empirically find a time asymmetry in their ability to learn natural language: a difference in the average log-perplexity when trying to predict the next token versus when trying to predict the previous one. This difference is at the same time subtle and very consistent across various modalities (language, model size, training time, ...). Theoretically, this is surprising: from an information-theoretic point of view, there should be no such difference. We provide a theoretical framework to explain how such an asymmetry can appear from sparsity and computational complexity considerations, and outline a number of perspectives opened by our results.
title Arrows of Time for Large Language Models
topic Machine Learning
Artificial Intelligence
Computation and Language
url https://arxiv.org/abs/2401.17505