{"id":"https://openalex.org/W7036096552","doi":null,"title":"Aligning Large Language Models During Inference Time","display_name":"Aligning Large Language Models During Inference Time","publication_year":2024,"publication_date":"2024-12-02","ids":{"openalex":"https://openalex.org/W7036096552"},"language":"en","primary_location":{"id":"pmh:oai:elib.dlr.de:210009","doi":null,"is_oa":true,"landing_page_url":"https://elib.dlr.de/210009/1/Aligning%20Large%20Language%20Models%20During%20Inference%20Time.pdf","pdf_url":"https://elib.dlr.de/210009/1/Aligning%20Large%20Language%20Models%20During%20Inference%20Time.pdf","source":{"id":"https://openalex.org/S4377196266","display_name":"elib (German Aerospace Center)","issn_l":null,"issn":null,"is_oa":false,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":"https://openalex.org/I2898391981","host_organization_name":"Deutsches Zentrum f\u00fcr Luft- und Raumfahrt e. V. (DLR)","host_organization_lineage":["https://openalex.org/I2898391981"],"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":"Hochschulschrift"},"volume":null,"type":"dissertation","is_traditional":true,"indexed_in":[],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://elib.dlr.de/210009/1/Aligning%20Large%20Language%20Models%20During%20Inference%20Time.pdf","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":null,"display_name":"Vogt, Julian","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":true,"raw_author_name":"Vogt, Julian","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":true,"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":true,"primary_topic":{"id":"https://openalex.org/T10028","display_name":"Topic Modeling","score":0.4293999969959259,"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":0.4293999969959259,"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":0.17499999701976776,"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/T10883","display_name":"Ethics and Social Impacts of AI","score":0.16380000114440918,"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/large-language-model-alignment","display_name":"large language model alignment","score":0.8709999918937683},{"id":"https://openalex.org/keywords/inference-time-alignment","display_name":"inference-time alignment","score":0.8450000286102295},{"id":"https://openalex.org/keywords/steering-vectors","display_name":"steering vectors","score":0.9570000171661377},{"id":"https://openalex.org/keywords/misalignment-detection","display_name":"misalignment detection","score":0.675000011920929},{"id":"https://openalex.org/keywords/few-shot-text-classification","display_name":"few-shot text classification","score":0.9309999942779541}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.6297000050544739},{"id":"https://openalex.org/C2776214188","wikidata":"https://www.wikidata.org/wiki/Q408386","display_name":"Inference","level":2,"score":0.5640000104904175},{"id":"https://openalex.org/C66322947","wikidata":"https://www.wikidata.org/wiki/Q11658","display_name":"Transformer","level":3,"score":0.4943999946117401},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.4244999885559082},{"id":"https://openalex.org/C99498987","wikidata":"https://www.wikidata.org/wiki/Q2210247","display_name":"Noise (video)","level":3,"score":0.3840000033378601},{"id":"https://openalex.org/C2780354894","wikidata":"https://www.wikidata.org/wiki/Q679300","display_name":"Steering wheel","level":2,"score":0.37290000915527344},{"id":"https://openalex.org/C137293760","wikidata":"https://www.wikidata.org/wiki/Q3621696","display_name":"Language model","level":2,"score":0.3449000120162964},{"id":"https://openalex.org/C195324797","wikidata":"https://www.wikidata.org/wiki/Q33742","display_name":"Natural language","level":2,"score":0.3336000144481659},{"id":"https://openalex.org/C2778572836","wikidata":"https://www.wikidata.org/wiki/Q380933","display_name":"Space (punctuation)","level":2,"score":0.3278999924659729},{"id":"https://openalex.org/C63479239","wikidata":"https://www.wikidata.org/wiki/Q7353546","display_name":"Robustness (evolution)","level":3,"score":0.2851000130176544},{"id":"https://openalex.org/C107457646","wikidata":"https://www.wikidata.org/wiki/Q207434","display_name":"Human\u2013computer interaction","level":1,"score":0.26579999923706055},{"id":"https://openalex.org/C83665646","wikidata":"https://www.wikidata.org/wiki/Q42139305","display_name":"Feature vector","level":2,"score":0.2653000056743622},{"id":"https://openalex.org/C12267149","wikidata":"https://www.wikidata.org/wiki/Q282453","display_name":"Support vector machine","level":2,"score":0.260699987411499},{"id":"https://openalex.org/C78639753","wikidata":"https://www.wikidata.org/wiki/Q3318160","display_name":"Behavioral modeling","level":2,"score":0.2597000002861023},{"id":"https://openalex.org/C143148639","wikidata":"https://www.wikidata.org/wiki/Q4876192","display_name":"Beam steering","level":3,"score":0.25870001316070557},{"id":"https://openalex.org/C2988790929","wikidata":"https://www.wikidata.org/wiki/Q125977","display_name":"Space vector","level":4,"score":0.2583000063896179},{"id":"https://openalex.org/C44154836","wikidata":"https://www.wikidata.org/wiki/Q45045","display_name":"Simulation","level":1,"score":0.2572000026702881},{"id":"https://openalex.org/C2778679013","wikidata":"https://www.wikidata.org/wiki/Q1054734","display_name":"Power steering","level":2,"score":0.257099986076355},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.2554999887943268},{"id":"https://openalex.org/C2777311342","wikidata":"https://www.wikidata.org/wiki/Q578801","display_name":"Shared space","level":3,"score":0.25110000371932983}],"mesh":[],"locations_count":1,"locations":[{"id":"pmh:oai:elib.dlr.de:210009","doi":null,"is_oa":true,"landing_page_url":"https://elib.dlr.de/210009/1/Aligning%20Large%20Language%20Models%20During%20Inference%20Time.pdf","pdf_url":"https://elib.dlr.de/210009/1/Aligning%20Large%20Language%20Models%20During%20Inference%20Time.pdf","source":{"id":"https://openalex.org/S4377196266","display_name":"elib (German Aerospace Center)","issn_l":null,"issn":null,"is_oa":false,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":"https://openalex.org/I2898391981","host_organization_name":"Deutsches Zentrum f\u00fcr Luft- und Raumfahrt e. V. (DLR)","host_organization_lineage":["https://openalex.org/I2898391981"],"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":"Hochschulschrift"}],"best_oa_location":{"id":"pmh:oai:elib.dlr.de:210009","doi":null,"is_oa":true,"landing_page_url":"https://elib.dlr.de/210009/1/Aligning%20Large%20Language%20Models%20During%20Inference%20Time.pdf","pdf_url":"https://elib.dlr.de/210009/1/Aligning%20Large%20Language%20Models%20During%20Inference%20Time.pdf","source":{"id":"https://openalex.org/S4377196266","display_name":"elib (German Aerospace Center)","issn_l":null,"issn":null,"is_oa":false,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":"https://openalex.org/I2898391981","host_organization_name":"Deutsches Zentrum f\u00fcr Luft- und Raumfahrt e. V. (DLR)","host_organization_lineage":["https://openalex.org/I2898391981"],"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":"Hochschulschrift"},"sustainable_development_goals":[],"study_designs":[],"awards":[],"funders":[{"id":"https://openalex.org/F4320321602","display_name":"Deutsches Zentrum f\u00fcr Luft- und Raumfahrt","ror":"https://ror.org/04bwf3e34"}],"has_content":{"grobid_xml":true,"pdf":true},"content_urls":{"pdf":"https://content.openalex.org/works/W7036096552.pdf","grobid_xml":"https://content.openalex.org/works/W7036096552.grobid-xml"},"referenced_works_count":65,"referenced_works":["https://openalex.org/W1614298861","https://openalex.org/W2081797125","https://openalex.org/W2176412452","https://openalex.org/W2468328197","https://openalex.org/W2525778437","https://openalex.org/W2747680751","https://openalex.org/W2789541106","https://openalex.org/W2804771341","https://openalex.org/W2883118930","https://openalex.org/W2885185669","https://openalex.org/W2922301133","https://openalex.org/W2950635152","https://openalex.org/W2994747431","https://openalex.org/W3015650676","https://openalex.org/W3017367042","https://openalex.org/W3027879771","https://openalex.org/W3035318851","https://openalex.org/W3037932933","https://openalex.org/W3089809702","https://openalex.org/W3102209588","https://openalex.org/W3128728164","https://openalex.org/W3144293453","https://openalex.org/W3201541172","https://openalex.org/W4308023630","https://openalex.org/W4311000453","https://openalex.org/W4312091446","https://openalex.org/W4327810158","https://openalex.org/W4377865076","https://openalex.org/W4386081218","https://openalex.org/W4387355345","https://openalex.org/W4387634876","https://openalex.org/W4387963693","https://openalex.org/W4387995257","https://openalex.org/W4388684774","https://openalex.org/W4389821460","https://openalex.org/W4390833718","https://openalex.org/W4390961118","https://openalex.org/W4391556798","https://openalex.org/W4392669749","https://openalex.org/W4392678154","https://openalex.org/W4392736799","https://openalex.org/W4395065622","https://openalex.org/W4399795335","https://openalex.org/W4400025319","https://openalex.org/W4401202605","https://openalex.org/W4402670277","https://openalex.org/W4402671016","https://openalex.org/W4402671231","https://openalex.org/W4403875167","https://openalex.org/W4404199641","https://openalex.org/W4404792953","https://openalex.org/W4406072474","https://openalex.org/W4409671508","https://openalex.org/W4412944437","https://openalex.org/W4415207023","https://openalex.org/W4415795446","https://openalex.org/W4415795733","https://openalex.org/W4415796496","https://openalex.org/W4415800641","https://openalex.org/W4416296080","https://openalex.org/W7103748068","https://openalex.org/W7117871696","https://openalex.org/W7133199456","https://openalex.org/W7133239970","https://openalex.org/W7134117706"],"related_works":["https://openalex.org/W4391157551","https://openalex.org/W4416022603","https://openalex.org/W7164031432","https://openalex.org/W4407131570","https://openalex.org/W7153005934","https://openalex.org/W7113243488","https://openalex.org/W4417101771","https://openalex.org/W4415193668","https://openalex.org/W7155244536","https://openalex.org/W4415796115"],"abstract_inverted_index":{"Large":[0],"language":[1,10],"models":[2],"have":[3,13,71],"led":[4],"to":[5,25,104,184,205,236,278,315,346],"significant":[6,221],"advancements":[7],"in":[8,63,111,165,177,194,220,239],"natural":[9],"processing":[11],"and":[12,45,59,89,96,102,118,139,156,207,217,245,267,295,308,311,355],"become":[14],"an":[15,73,260],"integral":[16],"part":[17],"of":[18,85,161,211,272,290,353],"everyday":[19],"life.":[20],"While":[21],"they":[22],"are":[23,122],"able":[24,183],"perform":[26],"various":[27],"tasks":[28],"with":[29,323],"increasing":[30],"accuracy,":[31],"sensitive":[32],"domains":[33],"such":[34],"as":[35],"healthcare":[36],"or":[37],"justice":[38],"place":[39],"high":[40,324],"demands":[41],"on":[42],"their":[43],"safety":[44],"reliability.":[46],"Models":[47],"that":[48,76,231,263,282],"do":[49],"not":[50,333],"follow":[51],"our":[52],"ethical":[53],"standards":[54],"can":[55],"produce":[56],"harmful":[57],"results":[58],"permanently":[60],"damage":[61],"trust":[62],"artificial":[64],"intelligence.":[65],"To":[66],"mitigate":[67],"this":[68],"risk,":[69],"we":[70,181],"developed":[72,230,296],"alignment":[74,126,188,251,266,352],"technique":[75,138],"operates":[77],"entirely":[78],"during":[79,92],"inference.":[80],"It":[81,302],"extracts":[82],"the":[83,93,119,125,145,148,178,187,191,195,198,212,240,250,265,268,287,291,304,312,317,337,341,351],"activations":[84],"a":[86,166,225,297,347,357],"few":[87],"positive":[88,155,307],"negative":[90,158,309],"examples":[91,159,310],"forward":[94,113],"pass,":[95],"then":[97],"uses":[98,232],"latent":[99,233],"space":[100,234],"arithmetic":[101,235],"post-processing":[103],"generate":[105],"effective":[106],"steering":[107,120,149,170,173,213,227,248,293],"vectors.":[108],"A":[109],"misalignment":[110,238],"subsequent":[112],"passes":[114],"is":[115,127,252],"automatically":[116],"detected":[117],"vectors":[121],"applied":[123,204],"until":[124,249],"restored.":[128,253],"We":[129,285],"started":[130],"by":[131,259],"implementing":[132],"Turner":[133],"et":[134],"al.'s":[135],"Activation":[136],"Addition":[137],"iteratively":[140],"improved":[141],"it":[142,331],"\\cite{activation-addition}.":[143],"In":[144,197],"first":[146],"iteration,":[147,200],"vector":[150,214],"was":[151,203,229,257],"obtained":[152],"from":[153],"50":[154,157,306],"instead":[160],"just":[162],"one,":[163],"resulting":[164,219],"nearly":[167],"bias-free":[168],"mean":[169],"vector.":[171],"By":[172],"over":[174,350],"multiple":[175],"layers":[176],"transformer":[179],"stack,":[180],"were":[182],"gradually":[185],"increase":[186],"without":[189],"overwriting":[190],"information":[192],"contained":[193],"embeddings.":[196],"third":[199],"Welch's":[201],"t-test":[202],"identify":[206],"eliminate":[208],"irrelevant":[209],"dimensions":[210],"containing":[215],"noise":[216],"bias,":[218],"performance":[222,270],"improvements.":[223],"Finally,":[224],"self-regulating":[226,292],"system":[228,294],"detect":[237],"embeddings":[241],"at":[242,319],"any":[243,320],"time":[244],"autonomously":[246],"starts":[247],"The":[254],"development":[255],"process":[256],"accompanied":[258],"evaluation":[261],"framework":[262],"quantified":[264],"associated":[269],"loss":[271],"each":[273],"modification.":[274],"This":[275],"allowed":[276],"us":[277],"adopt":[279],"only":[280],"those":[281],"provided":[283],"improvement.":[284],"extracted":[286],"detection":[288],"mechanism":[289],"token-wise":[298],"few-shot":[299],"text":[300],"classifier.":[301],"used":[303],"same":[305],"decoder-only":[313],"model":[314],"determine":[316],"sentiment":[318,328,338],"token":[321],"position":[322],"accuracy.":[325],"Unlike":[326],"scalar":[327],"analysis":[329],"models,":[330],"does":[332],"get":[334],"confused":[335],"when":[336],"changes":[339],"within":[340],"sentence.":[342],"Our":[343],"work":[344],"contributes":[345],"comprehensive":[348],"control":[349],"LLMs":[354],"represents":[356],"further":[358],"step":[359],"towards":[360],"safe":[361],"AI":[362],"models.":[363]},"counts_by_year":[],"updated_date":"2026-11-07T11:25:07.637180","created_date":"2025-10-10T00:00:00"}