Guardado en:
Detalles Bibliográficos
Autores principales: Brumley, Madeline, Kwon, Joe, Krueger, David, Krasheninnikov, Dmitrii, Anwar, Usman
Formato: Preprint
Publicado: 2024
Materias:
Acceso en línea:https://arxiv.org/abs/2411.07213
Etiquetas: Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
_version_ 1866917833984704512
author Brumley, Madeline
Kwon, Joe
Krueger, David
Krasheninnikov, Dmitrii
Anwar, Usman
author_facet Brumley, Madeline
Kwon, Joe
Krueger, David
Krasheninnikov, Dmitrii
Anwar, Usman
contents A key objective of interpretability research on large language models (LLMs) is to develop methods for robustly steering models toward desired behaviors. To this end, two distinct approaches to interpretability -- ``bottom-up" and ``top-down" -- have been presented, but there has been little quantitative comparison between them. We present a case study comparing the effectiveness of representative vector steering methods from each branch: function vectors (FV; arXiv:2310.15213), as a bottom-up method, and in-context vectors (ICV; arXiv:2311.06668) as a top-down method. While both aim to capture compact representations of broad in-context learning tasks, we find they are effective only on specific types of tasks: ICVs outperform FVs in behavioral shifting, whereas FVs excel in tasks requiring more precision. We discuss the implications for future evaluations of steering methods and for further research into top-down and bottom-up steering given these findings.
format Preprint
id arxiv_https___arxiv_org_abs_2411_07213
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle Comparing Bottom-Up and Top-Down Steering Approaches on In-Context Learning Tasks
Brumley, Madeline
Kwon, Joe
Krueger, David
Krasheninnikov, Dmitrii
Anwar, Usman
Machine Learning
A key objective of interpretability research on large language models (LLMs) is to develop methods for robustly steering models toward desired behaviors. To this end, two distinct approaches to interpretability -- ``bottom-up" and ``top-down" -- have been presented, but there has been little quantitative comparison between them. We present a case study comparing the effectiveness of representative vector steering methods from each branch: function vectors (FV; arXiv:2310.15213), as a bottom-up method, and in-context vectors (ICV; arXiv:2311.06668) as a top-down method. While both aim to capture compact representations of broad in-context learning tasks, we find they are effective only on specific types of tasks: ICVs outperform FVs in behavioral shifting, whereas FVs excel in tasks requiring more precision. We discuss the implications for future evaluations of steering methods and for further research into top-down and bottom-up steering given these findings.
title Comparing Bottom-Up and Top-Down Steering Approaches on In-Context Learning Tasks
topic Machine Learning
url https://arxiv.org/abs/2411.07213