Enregistré dans:
Détails bibliographiques
Auteurs principaux: Sun, Alan, Sun, Ethan, Shepard, Warren
Format: Preprint
Publié: 2024
Sujets:
Accès en ligne:https://arxiv.org/abs/2412.07386
Tags: Ajouter un tag
Pas de tags, Soyez le premier à ajouter un tag!
_version_ 1866915057094361088
author Sun, Alan
Sun, Ethan
Shepard, Warren
author_facet Sun, Alan
Sun, Ethan
Shepard, Warren
contents Zero-shot capabilities of large language models make them powerful tools for solving a range of tasks without explicit training. It remains unclear, however, how these models achieve such performance, or why they can zero-shot some tasks but not others. In this paper, we shed some light on this phenomenon by defining and investigating algorithmic stability in language models -- changes in problem-solving strategy employed by the model as a result of changes in task specification. We focus on a task where algorithmic stability is needed for generalization: two-operand arithmetic. Surprisingly, we find that Gemma-2-2b employs substantially different computational models on closely related subtasks, i.e. four-digit versus eight-digit addition. Our findings suggest that algorithmic instability may be a contributing factor to language models' poor zero-shot performance across certain logical reasoning tasks, as they struggle to abstract different problem-solving strategies and smoothly transition between them.
format Preprint
id arxiv_https___arxiv_org_abs_2412_07386
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle Algorithmic Phase Transitions in Language Models: A Mechanistic Case Study of Arithmetic
Sun, Alan
Sun, Ethan
Shepard, Warren
Computation and Language
Zero-shot capabilities of large language models make them powerful tools for solving a range of tasks without explicit training. It remains unclear, however, how these models achieve such performance, or why they can zero-shot some tasks but not others. In this paper, we shed some light on this phenomenon by defining and investigating algorithmic stability in language models -- changes in problem-solving strategy employed by the model as a result of changes in task specification. We focus on a task where algorithmic stability is needed for generalization: two-operand arithmetic. Surprisingly, we find that Gemma-2-2b employs substantially different computational models on closely related subtasks, i.e. four-digit versus eight-digit addition. Our findings suggest that algorithmic instability may be a contributing factor to language models' poor zero-shot performance across certain logical reasoning tasks, as they struggle to abstract different problem-solving strategies and smoothly transition between them.
title Algorithmic Phase Transitions in Language Models: A Mechanistic Case Study of Arithmetic
topic Computation and Language
url https://arxiv.org/abs/2412.07386