Comparing Bottom-Up and Top-Down Steering Approaches on In-Context Learning Tasks
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 litt...
Gespeichert in:
Veröffentlicht in: | arXiv.org 2024-11 |
---|---|
Hauptverfasser: | , , , , |
Format: | Artikel |
Sprache: | eng |
Schlagworte: | |
Online-Zugang: | Volltext |
Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
container_end_page | |
---|---|
container_issue | |
container_start_page | |
container_title | arXiv.org |
container_volume | |
creator | Brumley, Madeline Kwon, Joe Krueger, David Krasheninnikov, Dmitrii Usman Anwar |
description | 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 | Article |
fullrecord | <record><control><sourceid>proquest</sourceid><recordid>TN_cdi_proquest_journals_3127418680</recordid><sourceformat>XML</sourceformat><sourcesystem>PC</sourcesystem><sourcerecordid>3127418680</sourcerecordid><originalsourceid>FETCH-proquest_journals_31274186803</originalsourceid><addsrcrecordid>eNqNi0sKwjAUAIMgWLR3CLgOpEl_W62KghuxXZeg8VPtezFJ0eNbxQO4msXMDEggpIxYHgsxIqFzDedcpJlIEhmQXYGtUfYKZzpH77FllaEKjrREwxb4BLr3Wn_9zBiL6nDRjiLQDbACweuXp1utLHyKUrmbm5DhSd2dDn8ck-lqWRZr1t-PTjtfN9hZ6FUtI5HFUZ7mXP5XvQFFij6b</addsrcrecordid><sourcetype>Aggregation Database</sourcetype><iscdi>true</iscdi><recordtype>article</recordtype><pqid>3127418680</pqid></control><display><type>article</type><title>Comparing Bottom-Up and Top-Down Steering Approaches on In-Context Learning Tasks</title><source>Free E- Journals</source><creator>Brumley, Madeline ; Kwon, Joe ; Krueger, David ; Krasheninnikov, Dmitrii ; Usman Anwar</creator><creatorcontrib>Brumley, Madeline ; Kwon, Joe ; Krueger, David ; Krasheninnikov, Dmitrii ; Usman Anwar</creatorcontrib><description>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.</description><identifier>EISSN: 2331-8422</identifier><language>eng</language><publisher>Ithaca: Cornell University Library, arXiv.org</publisher><subject>Cognitive tasks ; Context ; Effectiveness ; Large language models ; Steering</subject><ispartof>arXiv.org, 2024-11</ispartof><rights>2024. This work is published under http://creativecommons.org/licenses/by/4.0/ (the “License”). Notwithstanding the ProQuest Terms and Conditions, you may use this content in accordance with the terms of the License.</rights><oa>free_for_read</oa><woscitedreferencessubscribed>false</woscitedreferencessubscribed></display><links><openurl>$$Topenurl_article</openurl><openurlfulltext>$$Topenurlfull_article</openurlfulltext><thumbnail>$$Tsyndetics_thumb_exl</thumbnail><link.rule.ids>780,784</link.rule.ids></links><search><creatorcontrib>Brumley, Madeline</creatorcontrib><creatorcontrib>Kwon, Joe</creatorcontrib><creatorcontrib>Krueger, David</creatorcontrib><creatorcontrib>Krasheninnikov, Dmitrii</creatorcontrib><creatorcontrib>Usman Anwar</creatorcontrib><title>Comparing Bottom-Up and Top-Down Steering Approaches on In-Context Learning Tasks</title><title>arXiv.org</title><description>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.</description><subject>Cognitive tasks</subject><subject>Context</subject><subject>Effectiveness</subject><subject>Large language models</subject><subject>Steering</subject><issn>2331-8422</issn><fulltext>true</fulltext><rsrctype>article</rsrctype><creationdate>2024</creationdate><recordtype>article</recordtype><sourceid>ABUWG</sourceid><sourceid>AFKRA</sourceid><sourceid>AZQEC</sourceid><sourceid>BENPR</sourceid><sourceid>CCPQU</sourceid><sourceid>DWQXO</sourceid><recordid>eNqNi0sKwjAUAIMgWLR3CLgOpEl_W62KghuxXZeg8VPtezFJ0eNbxQO4msXMDEggpIxYHgsxIqFzDedcpJlIEhmQXYGtUfYKZzpH77FllaEKjrREwxb4BLr3Wn_9zBiL6nDRjiLQDbACweuXp1utLHyKUrmbm5DhSd2dDn8ck-lqWRZr1t-PTjtfN9hZ6FUtI5HFUZ7mXP5XvQFFij6b</recordid><startdate>20241111</startdate><enddate>20241111</enddate><creator>Brumley, Madeline</creator><creator>Kwon, Joe</creator><creator>Krueger, David</creator><creator>Krasheninnikov, Dmitrii</creator><creator>Usman Anwar</creator><general>Cornell University Library, arXiv.org</general><scope>8FE</scope><scope>8FG</scope><scope>ABJCF</scope><scope>ABUWG</scope><scope>AFKRA</scope><scope>AZQEC</scope><scope>BENPR</scope><scope>BGLVJ</scope><scope>CCPQU</scope><scope>DWQXO</scope><scope>HCIFZ</scope><scope>L6V</scope><scope>M7S</scope><scope>PIMPY</scope><scope>PQEST</scope><scope>PQQKQ</scope><scope>PQUKI</scope><scope>PRINS</scope><scope>PTHSS</scope></search><sort><creationdate>20241111</creationdate><title>Comparing Bottom-Up and Top-Down Steering Approaches on In-Context Learning Tasks</title><author>Brumley, Madeline ; Kwon, Joe ; Krueger, David ; Krasheninnikov, Dmitrii ; Usman Anwar</author></sort><facets><frbrtype>5</frbrtype><frbrgroupid>cdi_FETCH-proquest_journals_31274186803</frbrgroupid><rsrctype>articles</rsrctype><prefilter>articles</prefilter><language>eng</language><creationdate>2024</creationdate><topic>Cognitive tasks</topic><topic>Context</topic><topic>Effectiveness</topic><topic>Large language models</topic><topic>Steering</topic><toplevel>online_resources</toplevel><creatorcontrib>Brumley, Madeline</creatorcontrib><creatorcontrib>Kwon, Joe</creatorcontrib><creatorcontrib>Krueger, David</creatorcontrib><creatorcontrib>Krasheninnikov, Dmitrii</creatorcontrib><creatorcontrib>Usman Anwar</creatorcontrib><collection>ProQuest SciTech Collection</collection><collection>ProQuest Technology Collection</collection><collection>Materials Science & Engineering Collection</collection><collection>ProQuest Central (Alumni Edition)</collection><collection>ProQuest Central UK/Ireland</collection><collection>ProQuest Central Essentials</collection><collection>ProQuest Central</collection><collection>Technology Collection</collection><collection>ProQuest One Community College</collection><collection>ProQuest Central Korea</collection><collection>SciTech Premium Collection</collection><collection>ProQuest Engineering Collection</collection><collection>Engineering Database</collection><collection>Access via ProQuest (Open Access)</collection><collection>ProQuest One Academic Eastern Edition (DO NOT USE)</collection><collection>ProQuest One Academic</collection><collection>ProQuest One Academic UKI Edition</collection><collection>ProQuest Central China</collection><collection>Engineering Collection</collection></facets><delivery><delcategory>Remote Search Resource</delcategory><fulltext>fulltext</fulltext></delivery><addata><au>Brumley, Madeline</au><au>Kwon, Joe</au><au>Krueger, David</au><au>Krasheninnikov, Dmitrii</au><au>Usman Anwar</au><format>book</format><genre>document</genre><ristype>GEN</ristype><atitle>Comparing Bottom-Up and Top-Down Steering Approaches on In-Context Learning Tasks</atitle><jtitle>arXiv.org</jtitle><date>2024-11-11</date><risdate>2024</risdate><eissn>2331-8422</eissn><abstract>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.</abstract><cop>Ithaca</cop><pub>Cornell University Library, arXiv.org</pub><oa>free_for_read</oa></addata></record> |
fulltext | fulltext |
identifier | EISSN: 2331-8422 |
ispartof | arXiv.org, 2024-11 |
issn | 2331-8422 |
language | eng |
recordid | cdi_proquest_journals_3127418680 |
source | Free E- Journals |
subjects | Cognitive tasks Context Effectiveness Large language models Steering |
title | Comparing Bottom-Up and Top-Down Steering Approaches on In-Context Learning Tasks |
url | https://sfx.bib-bvb.de/sfx_tum?ctx_ver=Z39.88-2004&ctx_enc=info:ofi/enc:UTF-8&ctx_tim=2024-12-24T22%3A16%3A31IST&url_ver=Z39.88-2004&url_ctx_fmt=infofi/fmt:kev:mtx:ctx&rfr_id=info:sid/primo.exlibrisgroup.com:primo3-Article-proquest&rft_val_fmt=info:ofi/fmt:kev:mtx:book&rft.genre=document&rft.atitle=Comparing%20Bottom-Up%20and%20Top-Down%20Steering%20Approaches%20on%20In-Context%20Learning%20Tasks&rft.jtitle=arXiv.org&rft.au=Brumley,%20Madeline&rft.date=2024-11-11&rft.eissn=2331-8422&rft_id=info:doi/&rft_dat=%3Cproquest%3E3127418680%3C/proquest%3E%3Curl%3E%3C/url%3E&disable_directlink=true&sfx.directlink=off&sfx.report_link=0&rft_id=info:oai/&rft_pqid=3127418680&rft_id=info:pmid/&rfr_iscdi=true |