Correcting Large Language Model Behavior via Influence Function
Recent advancements in AI alignment techniques have significantly improved the alignment of large language models (LLMs) with static human preferences. However, the dynamic nature of human preferences can render some prior training data outdated or even erroneous, ultimately causing LLMs to deviate...
Gespeichert in:
Veröffentlicht in: | arXiv.org 2024-12 |
---|---|
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 | Zhang, Han Zhang, Zhuo Zhang, Yi Zhai, Yuanzhao Peng, Hanyang Yu, Lei Yu, Yue Wang, Hui Liang, Bin Lin, Gui Xu, Ruifeng |
description | Recent advancements in AI alignment techniques have significantly improved the alignment of large language models (LLMs) with static human preferences. However, the dynamic nature of human preferences can render some prior training data outdated or even erroneous, ultimately causing LLMs to deviate from contemporary human preferences and societal norms. Existing methodologies, whether they involve the curation of new data for continual alignment or the manual correction of outdated data for re-alignment, demand costly human resources. To address this challenge, we propose a novel approach, Large Language Model Behavior Correction with Influence Function Recall and Post-Training (LANCET), which requires no human involvement. LANCET consists of two phases: (1) using influence functions to identify the training data that significantly impact undesirable model outputs, and (2) applying an Influence function-driven Bregman Optimization (IBO) technique to adjust the model's behavior based on these influence distributions. Our experiments demonstrate that LANCET effectively and efficiently correct inappropriate behaviors of LLMs. Furthermore, LANCET can outperform methods that rely on collecting human preferences, and it enhances the interpretability of learning human preferences within LLMs. |
format | Article |
fullrecord | <record><control><sourceid>proquest</sourceid><recordid>TN_cdi_proquest_journals_3148961360</recordid><sourceformat>XML</sourceformat><sourcesystem>PC</sourcesystem><sourcerecordid>3148961360</sourcerecordid><originalsourceid>FETCH-proquest_journals_31489613603</originalsourceid><addsrcrecordid>eNpjYuA0MjY21LUwMTLiYOAtLs4yMDAwMjM3MjU15mSwd84vKkpNLsnMS1fwSSxKTwWSeemliUCGb35Kao6CU2pGYllmfpFCWWaigmdeWk5pal5yqoJbaR5QU34eDwNrWmJOcSovlOZmUHZzDXH20C0oyi8sTS0uic_KLy3KA0rFGxuaWFiaGRqbGRgTpwoAyy04VQ</addsrcrecordid><sourcetype>Aggregation Database</sourcetype><iscdi>true</iscdi><recordtype>article</recordtype><pqid>3148961360</pqid></control><display><type>article</type><title>Correcting Large Language Model Behavior via Influence Function</title><source>Free E- Journals</source><creator>Zhang, Han ; Zhang, Zhuo ; Zhang, Yi ; Zhai, Yuanzhao ; Peng, Hanyang ; Yu, Lei ; Yu, Yue ; Wang, Hui ; Liang, Bin ; Lin, Gui ; Xu, Ruifeng</creator><creatorcontrib>Zhang, Han ; Zhang, Zhuo ; Zhang, Yi ; Zhai, Yuanzhao ; Peng, Hanyang ; Yu, Lei ; Yu, Yue ; Wang, Hui ; Liang, Bin ; Lin, Gui ; Xu, Ruifeng</creatorcontrib><description>Recent advancements in AI alignment techniques have significantly improved the alignment of large language models (LLMs) with static human preferences. However, the dynamic nature of human preferences can render some prior training data outdated or even erroneous, ultimately causing LLMs to deviate from contemporary human preferences and societal norms. Existing methodologies, whether they involve the curation of new data for continual alignment or the manual correction of outdated data for re-alignment, demand costly human resources. To address this challenge, we propose a novel approach, Large Language Model Behavior Correction with Influence Function Recall and Post-Training (LANCET), which requires no human involvement. LANCET consists of two phases: (1) using influence functions to identify the training data that significantly impact undesirable model outputs, and (2) applying an Influence function-driven Bregman Optimization (IBO) technique to adjust the model's behavior based on these influence distributions. Our experiments demonstrate that LANCET effectively and efficiently correct inappropriate behaviors of LLMs. Furthermore, LANCET can outperform methods that rely on collecting human preferences, and it enhances the interpretability of learning human preferences within LLMs.</description><identifier>EISSN: 2331-8422</identifier><language>eng</language><publisher>Ithaca: Cornell University Library, arXiv.org</publisher><subject>Alignment ; Influence functions ; Large language models</subject><ispartof>arXiv.org, 2024-12</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>Zhang, Han</creatorcontrib><creatorcontrib>Zhang, Zhuo</creatorcontrib><creatorcontrib>Zhang, Yi</creatorcontrib><creatorcontrib>Zhai, Yuanzhao</creatorcontrib><creatorcontrib>Peng, Hanyang</creatorcontrib><creatorcontrib>Yu, Lei</creatorcontrib><creatorcontrib>Yu, Yue</creatorcontrib><creatorcontrib>Wang, Hui</creatorcontrib><creatorcontrib>Liang, Bin</creatorcontrib><creatorcontrib>Lin, Gui</creatorcontrib><creatorcontrib>Xu, Ruifeng</creatorcontrib><title>Correcting Large Language Model Behavior via Influence Function</title><title>arXiv.org</title><description>Recent advancements in AI alignment techniques have significantly improved the alignment of large language models (LLMs) with static human preferences. However, the dynamic nature of human preferences can render some prior training data outdated or even erroneous, ultimately causing LLMs to deviate from contemporary human preferences and societal norms. Existing methodologies, whether they involve the curation of new data for continual alignment or the manual correction of outdated data for re-alignment, demand costly human resources. To address this challenge, we propose a novel approach, Large Language Model Behavior Correction with Influence Function Recall and Post-Training (LANCET), which requires no human involvement. LANCET consists of two phases: (1) using influence functions to identify the training data that significantly impact undesirable model outputs, and (2) applying an Influence function-driven Bregman Optimization (IBO) technique to adjust the model's behavior based on these influence distributions. Our experiments demonstrate that LANCET effectively and efficiently correct inappropriate behaviors of LLMs. Furthermore, LANCET can outperform methods that rely on collecting human preferences, and it enhances the interpretability of learning human preferences within LLMs.</description><subject>Alignment</subject><subject>Influence functions</subject><subject>Large language models</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>eNpjYuA0MjY21LUwMTLiYOAtLs4yMDAwMjM3MjU15mSwd84vKkpNLsnMS1fwSSxKTwWSeemliUCGb35Kao6CU2pGYllmfpFCWWaigmdeWk5pal5yqoJbaR5QU34eDwNrWmJOcSovlOZmUHZzDXH20C0oyi8sTS0uic_KLy3KA0rFGxuaWFiaGRqbGRgTpwoAyy04VQ</recordid><startdate>20241221</startdate><enddate>20241221</enddate><creator>Zhang, Han</creator><creator>Zhang, Zhuo</creator><creator>Zhang, Yi</creator><creator>Zhai, Yuanzhao</creator><creator>Peng, Hanyang</creator><creator>Yu, Lei</creator><creator>Yu, Yue</creator><creator>Wang, Hui</creator><creator>Liang, Bin</creator><creator>Lin, Gui</creator><creator>Xu, Ruifeng</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>20241221</creationdate><title>Correcting Large Language Model Behavior via Influence Function</title><author>Zhang, Han ; Zhang, Zhuo ; Zhang, Yi ; Zhai, Yuanzhao ; Peng, Hanyang ; Yu, Lei ; Yu, Yue ; Wang, Hui ; Liang, Bin ; Lin, Gui ; Xu, Ruifeng</author></sort><facets><frbrtype>5</frbrtype><frbrgroupid>cdi_FETCH-proquest_journals_31489613603</frbrgroupid><rsrctype>articles</rsrctype><prefilter>articles</prefilter><language>eng</language><creationdate>2024</creationdate><topic>Alignment</topic><topic>Influence functions</topic><topic>Large language models</topic><toplevel>online_resources</toplevel><creatorcontrib>Zhang, Han</creatorcontrib><creatorcontrib>Zhang, Zhuo</creatorcontrib><creatorcontrib>Zhang, Yi</creatorcontrib><creatorcontrib>Zhai, Yuanzhao</creatorcontrib><creatorcontrib>Peng, Hanyang</creatorcontrib><creatorcontrib>Yu, Lei</creatorcontrib><creatorcontrib>Yu, Yue</creatorcontrib><creatorcontrib>Wang, Hui</creatorcontrib><creatorcontrib>Liang, Bin</creatorcontrib><creatorcontrib>Lin, Gui</creatorcontrib><creatorcontrib>Xu, Ruifeng</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>Publicly Available Content Database</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>Zhang, Han</au><au>Zhang, Zhuo</au><au>Zhang, Yi</au><au>Zhai, Yuanzhao</au><au>Peng, Hanyang</au><au>Yu, Lei</au><au>Yu, Yue</au><au>Wang, Hui</au><au>Liang, Bin</au><au>Lin, Gui</au><au>Xu, Ruifeng</au><format>book</format><genre>document</genre><ristype>GEN</ristype><atitle>Correcting Large Language Model Behavior via Influence Function</atitle><jtitle>arXiv.org</jtitle><date>2024-12-21</date><risdate>2024</risdate><eissn>2331-8422</eissn><abstract>Recent advancements in AI alignment techniques have significantly improved the alignment of large language models (LLMs) with static human preferences. However, the dynamic nature of human preferences can render some prior training data outdated or even erroneous, ultimately causing LLMs to deviate from contemporary human preferences and societal norms. Existing methodologies, whether they involve the curation of new data for continual alignment or the manual correction of outdated data for re-alignment, demand costly human resources. To address this challenge, we propose a novel approach, Large Language Model Behavior Correction with Influence Function Recall and Post-Training (LANCET), which requires no human involvement. LANCET consists of two phases: (1) using influence functions to identify the training data that significantly impact undesirable model outputs, and (2) applying an Influence function-driven Bregman Optimization (IBO) technique to adjust the model's behavior based on these influence distributions. Our experiments demonstrate that LANCET effectively and efficiently correct inappropriate behaviors of LLMs. Furthermore, LANCET can outperform methods that rely on collecting human preferences, and it enhances the interpretability of learning human preferences within LLMs.</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, 2024-12 |
issn | 2331-8422 |
language | eng |
recordid | cdi_proquest_journals_3148961360 |
source | Free E- Journals |
subjects | Alignment Influence functions Large language models |
title | Correcting Large Language Model Behavior via Influence Function |
url | https://sfx.bib-bvb.de/sfx_tum?ctx_ver=Z39.88-2004&ctx_enc=info:ofi/enc:UTF-8&ctx_tim=2025-01-08T20%3A31%3A01IST&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=Correcting%20Large%20Language%20Model%20Behavior%20via%20Influence%20Function&rft.jtitle=arXiv.org&rft.au=Zhang,%20Han&rft.date=2024-12-21&rft.eissn=2331-8422&rft_id=info:doi/&rft_dat=%3Cproquest%3E3148961360%3C/proquest%3E%3Curl%3E%3C/url%3E&disable_directlink=true&sfx.directlink=off&sfx.report_link=0&rft_id=info:oai/&rft_pqid=3148961360&rft_id=info:pmid/&rfr_iscdi=true |