{"id":"https://openalex.org/W4414487768","doi":"https://doi.org/10.48550/arxiv.2508.12535","title":"CorrSteer: Generation-Time LLM Steering via Correlated Sparse Autoencoder Features","display_name":"CorrSteer: Generation-Time LLM Steering via Correlated Sparse Autoencoder Features","publication_year":2025,"publication_date":"2025-08-18","ids":{"openalex":"https://openalex.org/W4414487768","doi":"https://doi.org/10.48550/arxiv.2508.12535"},"language":"en","primary_location":{"id":"pmh:oai:arXiv.org:2508.12535","is_oa":true,"landing_page_url":"http://arxiv.org/abs/2508.12535","pdf_url":"https://arxiv.org/pdf/2508.12535","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,"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":null,"license_id":null,"version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":null},"type":"preprint","indexed_in":["arxiv","datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://arxiv.org/pdf/2508.12535","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5111062496","display_name":"Seonglae Cho","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Cho, Seonglae","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5046990142","display_name":"Zekun Wu","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Wu, Zekun","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5071702962","display_name":"Adriano Koshiyama","orcid":"https://orcid.org/0000-0001-7536-1503"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Koshiyama, Adriano","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]}],"institutions":[],"countries_distinct_count":0,"institutions_distinct_count":0,"corresponding_author_ids":[],"corresponding_institution_ids":[],"apc_list":null,"apc_paid":null,"fwci":null,"has_fulltext":false,"cited_by_count":0,"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/T12535","display_name":"Machine Learning and Data Classification","score":0.9090999960899353,"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/T12535","display_name":"Machine Learning and Data Classification","score":0.9090999960899353,"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/autoencoder","display_name":"Autoencoder","score":0.6464999914169312},{"id":"https://openalex.org/keywords/correctness","display_name":"Correctness","score":0.6219000220298767},{"id":"https://openalex.org/keywords/inference","display_name":"Inference","score":0.5649999976158142},{"id":"https://openalex.org/keywords/task","display_name":"Task (project management)","score":0.5648999810218811},{"id":"https://openalex.org/keywords/spurious-relationship","display_name":"Spurious relationship","score":0.560699999332428},{"id":"https://openalex.org/keywords/scalability","display_name":"Scalability","score":0.5145000219345093},{"id":"https://openalex.org/keywords/selection","display_name":"Selection (genetic algorithm)","score":0.4862000048160553},{"id":"https://openalex.org/keywords/pattern-recognition","display_name":"Pattern recognition (psychology)","score":0.4498000144958496}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.8037999868392944},{"id":"https://openalex.org/C101738243","wikidata":"https://www.wikidata.org/wiki/Q786435","display_name":"Autoencoder","level":3,"score":0.6464999914169312},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.6432999968528748},{"id":"https://openalex.org/C55439883","wikidata":"https://www.wikidata.org/wiki/Q360812","display_name":"Correctness","level":2,"score":0.6219000220298767},{"id":"https://openalex.org/C2776214188","wikidata":"https://www.wikidata.org/wiki/Q408386","display_name":"Inference","level":2,"score":0.5649999976158142},{"id":"https://openalex.org/C2780451532","wikidata":"https://www.wikidata.org/wiki/Q759676","display_name":"Task (project management)","level":2,"score":0.5648999810218811},{"id":"https://openalex.org/C97256817","wikidata":"https://www.wikidata.org/wiki/Q1462316","display_name":"Spurious relationship","level":2,"score":0.560699999332428},{"id":"https://openalex.org/C48044578","wikidata":"https://www.wikidata.org/wiki/Q727490","display_name":"Scalability","level":2,"score":0.5145000219345093},{"id":"https://openalex.org/C81917197","wikidata":"https://www.wikidata.org/wiki/Q628760","display_name":"Selection (genetic algorithm)","level":2,"score":0.4862000048160553},{"id":"https://openalex.org/C153180895","wikidata":"https://www.wikidata.org/wiki/Q7148389","display_name":"Pattern recognition (psychology)","level":2,"score":0.4498000144958496},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.4169999957084656},{"id":"https://openalex.org/C175154964","wikidata":"https://www.wikidata.org/wiki/Q380077","display_name":"Task analysis","level":3,"score":0.4115999937057495},{"id":"https://openalex.org/C198531522","wikidata":"https://www.wikidata.org/wiki/Q485146","display_name":"Sample (material)","level":2,"score":0.40700000524520874},{"id":"https://openalex.org/C2776401178","wikidata":"https://www.wikidata.org/wiki/Q12050496","display_name":"Feature (linguistics)","level":2,"score":0.35199999809265137},{"id":"https://openalex.org/C148483581","wikidata":"https://www.wikidata.org/wiki/Q446488","display_name":"Feature selection","level":2,"score":0.3246000111103058},{"id":"https://openalex.org/C137293760","wikidata":"https://www.wikidata.org/wiki/Q3621696","display_name":"Language model","level":2,"score":0.3179999887943268},{"id":"https://openalex.org/C124101348","wikidata":"https://www.wikidata.org/wiki/Q172491","display_name":"Data mining","level":1,"score":0.3034999966621399},{"id":"https://openalex.org/C52622490","wikidata":"https://www.wikidata.org/wiki/Q1026626","display_name":"Feature extraction","level":2,"score":0.2800000011920929},{"id":"https://openalex.org/C184337299","wikidata":"https://www.wikidata.org/wiki/Q1437428","display_name":"Semantics (computer science)","level":2,"score":0.2754000127315521},{"id":"https://openalex.org/C2778915421","wikidata":"https://www.wikidata.org/wiki/Q3643177","display_name":"Performance improvement","level":2,"score":0.2662000060081482},{"id":"https://openalex.org/C118505674","wikidata":"https://www.wikidata.org/wiki/Q42586063","display_name":"Encoder","level":2,"score":0.25110000371932983}],"mesh":[],"locations_count":3,"locations":[{"id":"pmh:oai:arXiv.org:2508.12535","is_oa":true,"landing_page_url":"http://arxiv.org/abs/2508.12535","pdf_url":"https://arxiv.org/pdf/2508.12535","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,"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":null,"license_id":null,"version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":null},{"id":"pmh:oai:arXiv.org:2508.12535","is_oa":true,"landing_page_url":"https://arxiv.org/abs/2508.12535","pdf_url":"https://arxiv.org/pdf/2508.12535","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,"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.2508.12535","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2508.12535","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,"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":null,"raw_source_name":null,"raw_type":"Preprint"}],"best_oa_location":{"id":"pmh:oai:arXiv.org:2508.12535","is_oa":true,"landing_page_url":"http://arxiv.org/abs/2508.12535","pdf_url":"https://arxiv.org/pdf/2508.12535","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,"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":null,"license_id":null,"version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":null},"sustainable_development_goals":[],"awards":[],"funders":[],"has_content":{"pdf":false,"grobid_xml":false},"content_urls":null,"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Sparse":[0],"Autoencoders":[1],"(SAEs)":[2],"can":[3],"extract":[4,63],"interpretable":[5],"features":[6,42,126],"from":[7,50,76],"large":[8,30],"language":[9,159],"models":[10],"(LLMs)":[11],"without":[12],"supervision.":[13],"However,":[14],"their":[15],"effectiveness":[16],"in":[17,109,119],"downstream":[18],"steering":[19,74,157],"tasks":[20],"is":[21],"limited":[22],"by":[23,43],"the":[24,80,137],"requirement":[25],"for":[26,154],"contrastive":[27],"datasets":[28],"or":[29],"activation":[31],"storage.":[32],"To":[33],"address":[34],"these":[35],"limitations,":[36],"we":[37],"propose":[38],"CorrSteer,":[39],"which":[40],"selects":[41],"correlating":[44],"sample":[45],"correctness":[46],"with":[47,112,121,132],"SAE":[48,156],"activations":[49,61],"generated":[51],"tokens":[52],"at":[53],"inference":[54],"time.":[55],"This":[56],"approach":[57,153],"uses":[58],"only":[59,122],"inference-time":[60],"to":[62],"more":[64],"relevant":[65],"features,":[66],"thereby":[67],"reducing":[68],"spurious":[69],"correlations.":[70],"It":[71],"also":[72],"obtains":[73],"coefficients":[75],"average":[77],"activations,":[78],"automating":[79],"entire":[81],"pipeline.":[82],"Our":[83,143],"method":[84],"shows":[85],"improved":[86],"task":[87],"performance":[88,111],"on":[89,98],"QA,":[90],"bias":[91],"mitigation,":[92],"jailbreaking":[93],"prevention,":[94],"and":[95,101,115,151],"reasoning":[96],"benchmarks":[97],"Gemma-2":[99],"2B":[100],"LLaMA-3.1":[102],"8B,":[103],"notably":[104],"achieving":[105],"a":[106,116],"+3.3%":[107],"improvement":[108,118],"MMLU":[110],"4000":[113],"samples":[114],"+27.2%":[117],"HarmBench":[120],"108":[123],"samples.":[124],"Selected":[125],"demonstrate":[127],"semantically":[128],"meaningful":[129],"patterns":[130],"aligned":[131],"each":[133],"task's":[134],"requirements,":[135],"revealing":[136],"underlying":[138],"capabilities":[139],"that":[140],"drive":[141],"performance.":[142],"work":[144],"establishes":[145],"correlation-based":[146],"selection":[147],"as":[148],"an":[149],"effective":[150],"scalable":[152],"automated":[155],"across":[158],"model":[160],"applications.":[161]},"counts_by_year":[],"updated_date":"2026-07-28T07:46:37.118299","created_date":"2025-09-25T00:00:00"}