{"id":"https://openalex.org/W4403578769","doi":"https://doi.org/10.48550/arxiv.2410.12877","title":"Improving Instruction-Following in Language Models through Activation Steering","display_name":"Improving Instruction-Following in Language Models through Activation Steering","publication_year":2024,"publication_date":"2024-10-15","ids":{"openalex":"https://openalex.org/W4403578769","doi":"https://doi.org/10.48550/arxiv.2410.12877"},"language":"en","primary_location":{"id":"pmh:oai:arXiv.org:2410.12877","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2410.12877","pdf_url":"https://arxiv.org/pdf/2410.12877","source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":"2331-8422","issn":["2331-8422"],"is_oa":true,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":"https://openalex.org/I205783295","host_organization_name":"Cornell University","host_organization_lineage":["https://openalex.org/I205783295"],"host_organization_lineage_names":[],"type":"repository"},"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},"volume":null,"type":"preprint","is_traditional":true,"indexed_in":["arxiv","datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://arxiv.org/pdf/2410.12877","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5067263653","display_name":"Alessandro Stolfo","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I35440088","display_name":"ETH Zurich","ror":"https://ror.org/05a28rw58","country_code":"CH","type":"education","lineage":["https://openalex.org/I2799323385","https://openalex.org/I35440088"]}],"countries":["CH"],"is_corresponding":false,"raw_author_name":"Stolfo, Alessandro","raw_affiliation_strings":["ETH Z\u00fcrich"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"ETH Z\u00fcrich","institution_ids":["https://openalex.org/I35440088"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5012491878","display_name":"Vidhisha Balachandran","orcid":"https://orcid.org/0009-0009-0465-0098","observed_orcids":["https://orcid.org/0009-0009-0465-0098"]},"institutions":[{"id":"https://openalex.org/I1290206253","display_name":"Microsoft (United States)","ror":"https://ror.org/00d0nc645","country_code":"US","type":"company","lineage":["https://openalex.org/I1290206253"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Balachandran, Vidhisha","raw_affiliation_strings":["Microsoft Research"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Microsoft Research","institution_ids":["https://openalex.org/I1290206253"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5103092774","display_name":"Safoora Yousefi","orcid":"https://orcid.org/0000-0001-7320-2998","observed_orcids":["https://orcid.org/0000-0001-7320-2998"]},"institutions":[{"id":"https://openalex.org/I1290206253","display_name":"Microsoft (United States)","ror":"https://ror.org/00d0nc645","country_code":"US","type":"company","lineage":["https://openalex.org/I1290206253"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Yousefi, Safoora","raw_affiliation_strings":["Microsoft Research"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Microsoft Research","institution_ids":["https://openalex.org/I1290206253"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5043228682","display_name":"Eric Horvitz","orcid":"https://orcid.org/0000-0002-8823-0614","observed_orcids":["https://orcid.org/0000-0002-8823-0614"]},"institutions":[{"id":"https://openalex.org/I1290206253","display_name":"Microsoft (United States)","ror":"https://ror.org/00d0nc645","country_code":"US","type":"company","lineage":["https://openalex.org/I1290206253"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Horvitz, Eric","raw_affiliation_strings":["Microsoft Research"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Microsoft Research","institution_ids":["https://openalex.org/I1290206253"]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5011998621","display_name":"Besmira Nushi","orcid":"https://orcid.org/0000-0002-7554-8586","observed_orcids":["https://orcid.org/0000-0002-7554-8586"]},"institutions":[{"id":"https://openalex.org/I1290206253","display_name":"Microsoft (United States)","ror":"https://ror.org/00d0nc645","country_code":"US","type":"company","lineage":["https://openalex.org/I1290206253"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Nushi, Besmira","raw_affiliation_strings":["Microsoft Research"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Microsoft Research","institution_ids":["https://openalex.org/I1290206253"]}]}],"institutions":[],"countries_distinct_count":2,"institutions_distinct_count":2,"corresponding_author_ids":[],"corresponding_institution_ids":[],"apc_list":null,"apc_paid":null,"fwci":null,"has_fulltext":true,"cited_by_count":71,"citation_normalized_percentile":null,"cited_by_percentile_year":null,"biblio":{"volume":null,"issue":null,"first_page":null,"last_page":null},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T10028","display_name":"Topic Modeling","score":0.885699987411499,"subfield":{"id":"https://openalex.org/subfields/1702","display_name":"Artificial Intelligence"},"field":{"id":"https://openalex.org/fields/17","display_name":"Computer Science"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}},"topics":[{"id":"https://openalex.org/T10028","display_name":"Topic Modeling","score":0.885699987411499,"subfield":{"id":"https://openalex.org/subfields/1702","display_name":"Artificial Intelligence"},"field":{"id":"https://openalex.org/fields/17","display_name":"Computer Science"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}},{"id":"https://openalex.org/T10181","display_name":"Natural Language Processing Techniques","score":0.03909999877214432,"subfield":{"id":"https://openalex.org/subfields/1702","display_name":"Artificial Intelligence"},"field":{"id":"https://openalex.org/fields/17","display_name":"Computer Science"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}},{"id":"https://openalex.org/T12026","display_name":"Explainable Artificial Intelligence (XAI)","score":0.01979999989271164,"subfield":{"id":"https://openalex.org/subfields/1702","display_name":"Artificial Intelligence"},"field":{"id":"https://openalex.org/fields/17","display_name":"Computer Science"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}}],"keywords":[{"id":"https://openalex.org/keywords/activation-steering","display_name":"activation steering","score":0.8980000019073486},{"id":"https://openalex.org/keywords/instruction-following","display_name":"instruction following","score":0.968999981880188},{"id":"https://openalex.org/keywords/language-models","display_name":"language models","score":0.9340000152587891},{"id":"https://openalex.org/keywords/inference-time-control","display_name":"inference-time control","score":0.6729999780654907},{"id":"https://openalex.org/keywords/knowledge-distillation","display_name":"knowledge distillation","score":0.7049999833106995},{"id":"https://openalex.org/keywords/instruction-tuned-models","display_name":"instruction-tuned models","score":0.6959999799728394},{"id":"https://openalex.org/keywords/language-generation","display_name":"language generation","score":0.640999972820282}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.44831445813179016}],"mesh":[],"locations_count":2,"locations":[{"id":"pmh:oai:arXiv.org:2410.12877","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2410.12877","pdf_url":"https://arxiv.org/pdf/2410.12877","source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":"2331-8422","issn":["2331-8422"],"is_oa":true,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":"https://openalex.org/I205783295","host_organization_name":"Cornell University","host_organization_lineage":["https://openalex.org/I205783295"],"host_organization_lineage_names":[],"type":"repository"},"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},{"id":"doi:10.48550/arxiv.2410.12877","doi":"https://doi.org/10.48550/arxiv.2410.12877","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2410.12877","pdf_url":null,"source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":"2331-8422","issn":["2331-8422"],"is_oa":true,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":"https://openalex.org/I205783295","host_organization_name":"Cornell University","host_organization_lineage":["https://openalex.org/I205783295"],"host_organization_lineage_names":[],"type":"repository"},"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":null,"is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"Preprint"}],"best_oa_location":{"id":"pmh:oai:arXiv.org:2410.12877","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2410.12877","pdf_url":"https://arxiv.org/pdf/2410.12877","source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":"2331-8422","issn":["2331-8422"],"is_oa":true,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":"https://openalex.org/I205783295","host_organization_name":"Cornell University","host_organization_lineage":["https://openalex.org/I205783295"],"host_organization_lineage_names":[],"type":"repository"},"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},"sustainable_development_goals":[],"study_designs":[],"awards":[],"funders":[{"id":"https://openalex.org/F4320308943","display_name":"Microsoft Research","ror":"https://ror.org/00d0nc645"}],"has_content":{"pdf":true,"grobid_xml":true},"content_urls":{"pdf":"https://content.openalex.org/works/W4403578769.pdf","grobid_xml":"https://content.openalex.org/works/W4403578769.grobid-xml"},"referenced_works_count":27,"referenced_works":["https://openalex.org/W2084413241","https://openalex.org/W2342249984","https://openalex.org/W3016682301","https://openalex.org/W3035965352","https://openalex.org/W3152515526","https://openalex.org/W3196731672","https://openalex.org/W3215698779","https://openalex.org/W4285149215","https://openalex.org/W4327810158","https://openalex.org/W4385374439","https://openalex.org/W4385572831","https://openalex.org/W4385572928","https://openalex.org/W4386081218","https://openalex.org/W4386290290","https://openalex.org/W4386566766","https://openalex.org/W4387355345","https://openalex.org/W4387561528","https://openalex.org/W4388718089","https://openalex.org/W4388787608","https://openalex.org/W4389518382","https://openalex.org/W4389518997","https://openalex.org/W4389519449","https://openalex.org/W4391505517","https://openalex.org/W4392120787","https://openalex.org/W4392736799","https://openalex.org/W4404752287","https://openalex.org/W4404784060"],"related_works":["https://openalex.org/W7134834324","https://openalex.org/W7170081662","https://openalex.org/W7160441609","https://openalex.org/W7128498199","https://openalex.org/W4407699659","https://openalex.org/W7223330018","https://openalex.org/W7160654272","https://openalex.org/W7162800591","https://openalex.org/W7140117179","https://openalex.org/W4416059721"],"abstract_inverted_index":{"The":[0],"ability":[1],"to":[2,34,57,69,98,101,109,141],"follow":[3,102],"instructions":[4,107,113,127],"is":[5],"crucial":[6],"for":[7,157],"numerous":[8],"real-world":[9],"applications":[10],"of":[11,16,121],"language":[12,29,161],"models.":[13,144],"In":[14],"pursuit":[15],"deeper":[17],"insights":[18],"and":[19,31,50,76,108,154,165],"more":[20],"powerful":[21],"capabilities,":[22],"we":[23,92,117,130],"derive":[24],"instruction-specific":[25],"vector":[26],"representations":[27],"from":[28],"models":[30,36,89,100,138],"use":[32,94],"them":[33],"steer":[35],"accordingly.":[37],"These":[38],"vectors":[39,97,134],"are":[40,114,167],"computed":[41,135],"as":[42,72],"the":[43,95,119],"difference":[44],"in":[45,160],"activations":[46],"between":[47],"inputs":[48],"with":[49],"without":[51,105],"instructions,":[52],"enabling":[53],"a":[54,152],"modular":[55],"approach":[56,156],"activation":[58,96,122,149],"steering.":[59],"We":[60],"demonstrate":[61,90,131,147],"how":[62,91],"this":[63],"method":[64],"can":[65,93,139],"enhance":[66,110],"model":[67],"adherence":[68],"constraints":[70,103],"such":[71],"output":[73],"format,":[74],"length,":[75],"word":[77],"inclusion,":[78],"providing":[79],"inference-time":[80],"control":[81,159],"over":[82],"instruction":[83],"following.":[84],"Our":[85,145,163],"experiments":[86],"across":[87],"four":[88],"guide":[99],"even":[104],"explicit":[106],"performance":[111],"when":[112],"present.":[115],"Additionally,":[116],"explore":[118],"compositionality":[120],"steering,":[123],"successfully":[124],"applying":[125],"multiple":[126],"simultaneously.":[128],"Finally,":[129],"that":[132,148],"steering":[133,150],"on":[136],"instruction-tuned":[137],"transfer":[140],"improve":[142],"base":[143],"findings":[146],"offers":[151],"practical":[153],"scalable":[155],"fine-grained":[158],"generation.":[162],"code":[164],"data":[166],"available":[168],"at":[169],"https://github.com/microsoft/llm-steer-instruct.":[170]},"counts_by_year":[{"year":2026,"cited_by_count":49},{"year":2025,"cited_by_count":21},{"year":2024,"cited_by_count":1}],"updated_date":"2026-11-07T11:25:07.637180","created_date":"2024-10-20T00:00:00"}