{"id":"https://openalex.org/W7160838046","doi":"https://doi.org/10.48550/arxiv.2605.07631","title":"Inference Time Causal Probing in LLMs","display_name":"Inference Time Causal Probing in LLMs","publication_year":2026,"publication_date":"2026-05-08","ids":{"openalex":"https://openalex.org/W7160838046","doi":"https://doi.org/10.48550/arxiv.2605.07631"},"language":null,"primary_location":{"id":"doi:10.48550/arxiv.2605.07631","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2605.07631","pdf_url":null,"source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":null,"issn":null,"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":false,"raw_source_name":null,"raw_type":"Preprint"},"type":"preprint","indexed_in":["datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://doi.org/10.48550/arxiv.2605.07631","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5108489130","display_name":"Sadegh Khorasani","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Khorasani, Sadegh","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5072294850","display_name":"Saber Salehkaleybar","orcid":"https://orcid.org/0000-0003-3934-9931"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Salehkaleybar, Saber","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5012789968","display_name":"Negar Kiyavash","orcid":"https://orcid.org/0000-0002-8545-7709"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Kiyavash, Negar","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5014508720","display_name":"Matthias Grossglauser","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Grossglauser, Matthias","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/T10028","display_name":"Topic Modeling","score":0.4697999954223633,"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":0.4697999954223633,"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/T10775","display_name":"Generative Adversarial Networks and Image Synthesis","score":0.08619999885559082,"subfield":{"id":"https://openalex.org/subfields/1707","display_name":"Computer Vision and Pattern Recognition"},"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/T11303","display_name":"Bayesian Modeling and Causal Inference","score":0.05130000039935112,"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/softmax-function","display_name":"Softmax function","score":0.671500027179718},{"id":"https://openalex.org/keywords/inference","display_name":"Inference","score":0.5613999962806702},{"id":"https://openalex.org/keywords/causal-inference","display_name":"Causal inference","score":0.5485000014305115},{"id":"https://openalex.org/keywords/margin","display_name":"Margin (machine learning)","score":0.5317999720573425},{"id":"https://openalex.org/keywords/measure","display_name":"Measure (data warehouse)","score":0.5306000113487244},{"id":"https://openalex.org/keywords/property","display_name":"Property (philosophy)","score":0.5303999781608582},{"id":"https://openalex.org/keywords/task","display_name":"Task (project management)","score":0.5291000008583069},{"id":"https://openalex.org/keywords/discriminator","display_name":"Discriminator","score":0.44279998540878296},{"id":"https://openalex.org/keywords/security-token","display_name":"Security token","score":0.4318000078201294}],"concepts":[{"id":"https://openalex.org/C188441871","wikidata":"https://www.wikidata.org/wiki/Q7554146","display_name":"Softmax function","level":3,"score":0.671500027179718},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.5695000290870667},{"id":"https://openalex.org/C2776214188","wikidata":"https://www.wikidata.org/wiki/Q408386","display_name":"Inference","level":2,"score":0.5613999962806702},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.5486000180244446},{"id":"https://openalex.org/C158600405","wikidata":"https://www.wikidata.org/wiki/Q5054566","display_name":"Causal inference","level":2,"score":0.5485000014305115},{"id":"https://openalex.org/C774472","wikidata":"https://www.wikidata.org/wiki/Q6760393","display_name":"Margin (machine learning)","level":2,"score":0.5317999720573425},{"id":"https://openalex.org/C2780009758","wikidata":"https://www.wikidata.org/wiki/Q6804172","display_name":"Measure (data warehouse)","level":2,"score":0.5306000113487244},{"id":"https://openalex.org/C189950617","wikidata":"https://www.wikidata.org/wiki/Q937228","display_name":"Property (philosophy)","level":2,"score":0.5303999781608582},{"id":"https://openalex.org/C2780451532","wikidata":"https://www.wikidata.org/wiki/Q759676","display_name":"Task (project management)","level":2,"score":0.5291000008583069},{"id":"https://openalex.org/C2779803651","wikidata":"https://www.wikidata.org/wiki/Q5282088","display_name":"Discriminator","level":3,"score":0.44279998540878296},{"id":"https://openalex.org/C48145219","wikidata":"https://www.wikidata.org/wiki/Q1335365","display_name":"Security token","level":2,"score":0.4318000078201294},{"id":"https://openalex.org/C95623464","wikidata":"https://www.wikidata.org/wiki/Q1096149","display_name":"Classifier (UML)","level":2,"score":0.4004000127315521},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.38920000195503235},{"id":"https://openalex.org/C11671645","wikidata":"https://www.wikidata.org/wiki/Q5054567","display_name":"Causal model","level":2,"score":0.3758000135421753},{"id":"https://openalex.org/C43214815","wikidata":"https://www.wikidata.org/wiki/Q7310987","display_name":"Reliability (semiconductor)","level":3,"score":0.35519999265670776},{"id":"https://openalex.org/C167966045","wikidata":"https://www.wikidata.org/wiki/Q5532625","display_name":"Generative model","level":3,"score":0.349700003862381},{"id":"https://openalex.org/C49937458","wikidata":"https://www.wikidata.org/wiki/Q2599292","display_name":"Probabilistic logic","level":2,"score":0.34299999475479126},{"id":"https://openalex.org/C2776502983","wikidata":"https://www.wikidata.org/wiki/Q690182","display_name":"Contrast (vision)","level":2,"score":0.3343999981880188},{"id":"https://openalex.org/C17231256","wikidata":"https://www.wikidata.org/wiki/Q5156540","display_name":"Completeness (order theory)","level":2,"score":0.32409998774528503},{"id":"https://openalex.org/C197640229","wikidata":"https://www.wikidata.org/wiki/Q2534066","display_name":"Predictability","level":2,"score":0.3077000081539154},{"id":"https://openalex.org/C88626702","wikidata":"https://www.wikidata.org/wiki/Q1128903","display_name":"Continuation","level":2,"score":0.2883000075817108},{"id":"https://openalex.org/C149782125","wikidata":"https://www.wikidata.org/wiki/Q160039","display_name":"Econometrics","level":1,"score":0.28630000352859497},{"id":"https://openalex.org/C2779662365","wikidata":"https://www.wikidata.org/wiki/Q5416694","display_name":"Event (particle physics)","level":2,"score":0.28519999980926514},{"id":"https://openalex.org/C33923547","wikidata":"https://www.wikidata.org/wiki/Q395","display_name":"Mathematics","level":0,"score":0.28369998931884766},{"id":"https://openalex.org/C204321447","wikidata":"https://www.wikidata.org/wiki/Q30642","display_name":"Natural language processing","level":1,"score":0.2806999981403351},{"id":"https://openalex.org/C137002209","wikidata":"https://www.wikidata.org/wiki/Q898521","display_name":"Hidden variable theory","level":3,"score":0.2800000011920929},{"id":"https://openalex.org/C48103436","wikidata":"https://www.wikidata.org/wiki/Q599031","display_name":"State (computer science)","level":2,"score":0.2718000113964081},{"id":"https://openalex.org/C15744967","wikidata":"https://www.wikidata.org/wiki/Q9418","display_name":"Psychology","level":0,"score":0.2655999958515167},{"id":"https://openalex.org/C2775924081","wikidata":"https://www.wikidata.org/wiki/Q55608371","display_name":"Control (management)","level":2,"score":0.2653000056743622},{"id":"https://openalex.org/C23224414","wikidata":"https://www.wikidata.org/wiki/Q176769","display_name":"Hidden Markov model","level":2,"score":0.262800008058548},{"id":"https://openalex.org/C63479239","wikidata":"https://www.wikidata.org/wiki/Q7353546","display_name":"Robustness (evolution)","level":3,"score":0.25929999351501465}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.48550/arxiv.2605.07631","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2605.07631","pdf_url":null,"source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":null,"issn":null,"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":"doi:10.48550/arxiv.2605.07631","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2605.07631","pdf_url":null,"source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":null,"issn":null,"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":false,"raw_source_name":null,"raw_type":"Preprint"},"sustainable_development_goals":[{"display_name":"Peace, Justice and strong institutions","id":"https://metadata.un.org/sdg/16","score":0.5133458375930786}],"awards":[],"funders":[],"has_content":{"pdf":false,"grobid_xml":false},"content_urls":null,"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Causal":[0],"probing":[1],"methods":[2,183],"aim":[3],"to":[4,50],"test":[5],"and":[6,56,158,165,189,195],"control":[7],"how":[8],"internal":[9],"representations":[10],"influence":[11],"the":[12,48,60,81,92,102,122,126,132,152,185,190],"behavior":[13],"of":[14,94,101,134,174],"generative":[15],"models.":[16],"In":[17],"causal":[18],"probing,":[19],"an":[20,42,171],"intervention":[21],"modifies":[22],"hidden":[23,78,128],"states":[24,79],"so":[25,130],"that":[26,75,90,100,119,131],"a":[27,31,51,71,87,95,112],"property":[28,154],"takes":[29],"on":[30,106,184],"different":[32],"value.":[33],"Most":[34],"existing":[35],"approaches":[36],"define":[37],"such":[38],"interventions":[39],"by":[40],"training":[41],"auxiliary":[43],"probe":[44,107],"classifier,":[45],"which":[46],"ties":[47],"method":[49],"specific":[52],"task":[53],"or":[54],"model":[55],"risks":[57],"misalignment":[58],"with":[59],"model's":[61,82],"predictive":[62],"geometry.":[63],"We":[64,109],"propose":[65],"Hidden-state":[66],"Driven":[67],"Margin":[68],"Intervention":[69],"(HDMI),":[70],"probe-free,":[72],"gradient-based":[73],"technique":[74],"directly":[76],"steers":[77],"using":[80],"native":[83],"output.":[84],"HDMI":[85,176],"applies":[86],"margin":[88],"objective":[89],"increases":[91,137],"probability":[93],"target":[96],"continuation":[97],"while":[98,142],"decreasing":[99],"source,":[103],"without":[104],"relying":[105],"classifiers.":[108],"further":[110],"introduce":[111],"lookahead":[113],"variant":[114],"(LA-HDMI)":[115],"for":[116],"text":[117],"editing":[118],"backpropagates":[120],"through":[121],"softmax":[123],"embeddings,":[124],"modifying":[125],"current":[127],"state":[129],"likelihood":[133],"user-specified":[135],"tokens":[136],"in":[138],"next":[139],"token":[140],"generations":[141],"preserving":[143],"fluency.":[144],"To":[145],"evaluate":[146],"interventions,":[147],"we":[148],"measure":[149,173],"completeness":[150],"(whether":[151,160],"targeted":[153],"changes":[155],"as":[156,170],"intended)":[157],"selectivity":[159],"unrelated":[161],"properties":[162],"are":[163],"preserved),":[164],"report":[166],"their":[167],"harmonic":[168],"mean":[169],"overall":[172],"reliability.":[175],"consistently":[177],"achieves":[178],"higher":[179],"reliability":[180],"than":[181],"prior":[182],"LGD":[186],"agreement":[187],"corpus":[188],"CausalGym":[191],"benchmark,":[192],"across":[193],"Meta-Llama-3-8B-Instruct,":[194],"Pythia-70M.":[196]},"counts_by_year":[],"updated_date":"2026-07-01T06:00:48.157686","created_date":"2026-05-12T00:00:00"}