{"id":"https://openalex.org/W4415232563","doi":"https://doi.org/10.18653/v1/2026.acl-long.2159","title":"GrAInS: Gradient-based Attribution for Inference-Time Steering of LLMs and VLMs","display_name":"GrAInS: Gradient-based Attribution for Inference-Time Steering of LLMs and VLMs","publication_year":2026,"publication_date":"2026-01-01","ids":{"openalex":"https://openalex.org/W4415232563","doi":"https://doi.org/10.18653/v1/2026.acl-long.2159"},"language":"en","primary_location":{"id":"doi:10.18653/v1/2026.acl-long.2159","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.acl-long.2159","pdf_url":null,"source":null,"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)","raw_type":"proceedings-article"},"type":"conference-paper","indexed_in":["arxiv","crossref","datacite"],"open_access":{"is_oa":true,"oa_status":"gold","oa_url":"https://doi.org/10.18653/v1/2026.acl-long.2159","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5101928156","display_name":"Duy Lap Nguyen","orcid":"https://orcid.org/0009-0001-4239-9174"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Duy Nguyen","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5039920828","display_name":"Archiki Prasad","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Archiki Prasad","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5038276570","display_name":"Elias Stengel-Eskin","orcid":"https://orcid.org/0000-0002-6689-505X"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Elias Stengel-Eskin","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5089199686","display_name":"Mohit Bansal","orcid":"https://orcid.org/0000-0001-5858-9944"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Mohit Bansal","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":43.6753,"has_fulltext":false,"cited_by_count":1,"citation_normalized_percentile":{"value":0.9832389,"is_in_top_1_percent":false,"is_in_top_10_percent":true},"cited_by_percentile_year":{"min":94,"max":97},"biblio":{"volume":null,"issue":null,"first_page":"46523","last_page":"46543"},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T14011","display_name":"Elevator Systems and Control","score":0.914900004863739,"subfield":{"id":"https://openalex.org/subfields/2207","display_name":"Control and Systems Engineering"},"field":{"id":"https://openalex.org/fields/22","display_name":"Engineering"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}},"topics":[{"id":"https://openalex.org/T14011","display_name":"Elevator Systems and Control","score":0.914900004863739,"subfield":{"id":"https://openalex.org/subfields/2207","display_name":"Control and Systems Engineering"},"field":{"id":"https://openalex.org/fields/22","display_name":"Engineering"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}}],"keywords":[{"id":"https://openalex.org/keywords/attribution","display_name":"Attribution","score":0.5853999853134155},{"id":"https://openalex.org/keywords/leverage","display_name":"Leverage (statistics)","score":0.5230000019073486},{"id":"https://openalex.org/keywords/language-model","display_name":"Language model","score":0.498199999332428},{"id":"https://openalex.org/keywords/fluency","display_name":"Fluency","score":0.48489999771118164},{"id":"https://openalex.org/keywords/modular-design","display_name":"Modular design","score":0.48190000653266907},{"id":"https://openalex.org/keywords/construct","display_name":"Construct (python library)","score":0.38830000162124634},{"id":"https://openalex.org/keywords/normative","display_name":"Normative","score":0.3564999997615814},{"id":"https://openalex.org/keywords/control","display_name":"Control (management)","score":0.35359999537467957}],"concepts":[{"id":"https://openalex.org/C143299363","wikidata":"https://www.wikidata.org/wiki/Q900584","display_name":"Attribution","level":2,"score":0.5853999853134155},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.5752999782562256},{"id":"https://openalex.org/C153083717","wikidata":"https://www.wikidata.org/wiki/Q6535263","display_name":"Leverage (statistics)","level":2,"score":0.5230000019073486},{"id":"https://openalex.org/C137293760","wikidata":"https://www.wikidata.org/wiki/Q3621696","display_name":"Language model","level":2,"score":0.498199999332428},{"id":"https://openalex.org/C2777413886","wikidata":"https://www.wikidata.org/wiki/Q3276013","display_name":"Fluency","level":2,"score":0.48489999771118164},{"id":"https://openalex.org/C180747234","wikidata":"https://www.wikidata.org/wiki/Q23373","display_name":"Cognitive psychology","level":1,"score":0.4846000075340271},{"id":"https://openalex.org/C101468663","wikidata":"https://www.wikidata.org/wiki/Q1620158","display_name":"Modular design","level":2,"score":0.48190000653266907},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.4772000014781952},{"id":"https://openalex.org/C15744967","wikidata":"https://www.wikidata.org/wiki/Q9418","display_name":"Psychology","level":0,"score":0.3946000039577484},{"id":"https://openalex.org/C2780801425","wikidata":"https://www.wikidata.org/wiki/Q5164392","display_name":"Construct (python library)","level":2,"score":0.38830000162124634},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.36010000109672546},{"id":"https://openalex.org/C44725695","wikidata":"https://www.wikidata.org/wiki/Q288156","display_name":"Normative","level":2,"score":0.3564999997615814},{"id":"https://openalex.org/C2775924081","wikidata":"https://www.wikidata.org/wiki/Q55608371","display_name":"Control (management)","level":2,"score":0.35359999537467957},{"id":"https://openalex.org/C66322947","wikidata":"https://www.wikidata.org/wiki/Q11658","display_name":"Transformer","level":3,"score":0.3531999886035919},{"id":"https://openalex.org/C2778712577","wikidata":"https://www.wikidata.org/wiki/Q3505966","display_name":"Retraining","level":2,"score":0.35010001063346863},{"id":"https://openalex.org/C204321447","wikidata":"https://www.wikidata.org/wiki/Q30642","display_name":"Natural language processing","level":1,"score":0.3476000130176544},{"id":"https://openalex.org/C63479239","wikidata":"https://www.wikidata.org/wiki/Q7353546","display_name":"Robustness (evolution)","level":3,"score":0.33169999718666077},{"id":"https://openalex.org/C11671645","wikidata":"https://www.wikidata.org/wiki/Q5054567","display_name":"Causal model","level":2,"score":0.3068000078201294},{"id":"https://openalex.org/C511192102","wikidata":"https://www.wikidata.org/wiki/Q5156948","display_name":"Comprehension","level":2,"score":0.2971000075340271},{"id":"https://openalex.org/C2776809875","wikidata":"https://www.wikidata.org/wiki/Q1375963","display_name":"Converse","level":2,"score":0.29589998722076416},{"id":"https://openalex.org/C2780665704","wikidata":"https://www.wikidata.org/wiki/Q959298","display_name":"Intervention (counseling)","level":2,"score":0.2849999964237213},{"id":"https://openalex.org/C158600405","wikidata":"https://www.wikidata.org/wiki/Q5054566","display_name":"Causal inference","level":2,"score":0.2653999924659729},{"id":"https://openalex.org/C2779918689","wikidata":"https://www.wikidata.org/wiki/Q3771842","display_name":"Stimulus (psychology)","level":2,"score":0.2653000056743622},{"id":"https://openalex.org/C165064840","wikidata":"https://www.wikidata.org/wiki/Q1321061","display_name":"Matching (statistics)","level":2,"score":0.259799987077713},{"id":"https://openalex.org/C2780791683","wikidata":"https://www.wikidata.org/wiki/Q846785","display_name":"Action (physics)","level":2,"score":0.25839999318122864},{"id":"https://openalex.org/C77805123","wikidata":"https://www.wikidata.org/wiki/Q161272","display_name":"Social psychology","level":1,"score":0.2563999891281128}],"mesh":[],"locations_count":3,"locations":[{"id":"doi:10.18653/v1/2026.acl-long.2159","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.acl-long.2159","pdf_url":null,"source":null,"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)","raw_type":"proceedings-article"},{"id":"pmh:oai:arXiv.org:2507.18043","is_oa":true,"landing_page_url":"https://arxiv.org/abs/2507.18043","pdf_url":"https://arxiv.org/pdf/2507.18043","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":"text"},{"id":"doi:10.48550/arxiv.2507.18043","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2507.18043","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":null,"license_id":null,"version":null,"is_accepted":false,"is_published":null,"raw_source_name":null,"raw_type":"Preprint"}],"best_oa_location":{"id":"doi:10.18653/v1/2026.acl-long.2159","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.acl-long.2159","pdf_url":null,"source":null,"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)","raw_type":"proceedings-article"},"sustainable_development_goals":[],"awards":[],"funders":[],"has_content":{"grobid_xml":false,"pdf":false},"content_urls":null,"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Inference-time":[0],"steering":[1,76,125,180],"methods":[2],"offer":[3],"a":[4,184],"lightweight":[5],"alternative":[6],"to":[7,48,96,112,122,133,153,199],"fine-tuning":[8,177],"large":[9],"language":[10],"models":[11,15,85],"(LLMs)":[12],"and":[13,46,62,83,86,105,150,161,178,203,218],"vision-language":[14,84],"(VLMs)":[16],"by":[17,146,210],"modifying":[18],"internal":[19],"activations":[20,141,152],"at":[21,142],"test":[22],"time":[23],"without":[24,167],"updating":[25],"model":[26,165],"weights.":[27],"However,":[28],"most":[29,100],"existing":[30,179],"approaches":[31],"rely":[32],"on":[33,109,188,195,208],"fixed,":[34],"global":[35],"intervention":[36],"vectors,":[37],"overlook":[38],"the":[39,53,98,215],"causal":[40],"influence":[41],"of":[42],"individual":[43],"input":[44],"tokens,":[45,102],"fail":[47],"leverage":[49],"informative":[50],"gradients":[51],"from":[52,131,197],"model's":[54,216],"logits,":[55],"particularly":[56],"in":[57],"multimodal":[58],"settings":[59],"where":[60],"visual":[61],"textual":[63],"inputs":[64],"contribute":[65],"unevenly.":[66],"To":[67],"address":[68],"these":[69],"limitations,":[70],"we":[71],"introduce":[72],"GrAInS,":[73],"an":[74],"inference-time":[75],"approach":[77],"that":[78,127],"operates":[79],"across":[80],"both":[81,103,176],"language-only":[82],"tasks.":[87],"GrAInS":[88,138,173],"uses":[89],"contrastive,":[90],"gradient-based":[91],"attribution":[92,148],"via":[93],"Integrated":[94],"Gradients":[95],"identify":[97],"top-k":[99],"influential":[101],"positively":[104],"negatively":[106],"attributed":[107],"based":[108],"their":[110],"contribution":[111],"preferred":[113],"versus":[114],"dispreferred":[115],"outputs.":[116],"These":[117],"tokens":[118],"are":[119],"then":[120],"used":[121],"construct":[123],"directional":[124],"vectors":[126],"capture":[128],"semantic":[129],"shifts":[130],"undesirable":[132],"desirable":[134],"behavior.":[135],"During":[136],"inference,":[137],"adjusts":[139],"hidden":[140],"transformer":[143],"layers":[144],"guided":[145],"token-level":[147],"signals,":[149],"normalizes":[151],"preserve":[154],"representational":[155],"scale.":[156],"This":[157],"enables":[158],"fine-grained,":[159],"interpretable,":[160],"modular":[162],"control":[163],"over":[164],"behavior,":[166],"retraining":[168],"or":[169],"auxiliary":[170],"supervision.":[171],"Empirically,":[172],"consistently":[174],"outperforms":[175],"baselines:":[181],"it":[182],"achieves":[183],"13.22%":[185],"accuracy":[186],"gain":[187],"TruthfulQA":[189],"using":[190],"Llama-3.1-8B,":[191],"reduces":[192],"hallucination":[193],"rates":[194,207],"MMHal-Bench":[196],"0.624":[198],"0.514":[200],"with":[201],"LLaVA-1.6-7B,":[202],"improves":[204],"alignment":[205],"win":[206],"SPA-VL":[209],"8.11%,":[211],"all":[212],"while":[213],"preserving":[214],"fluency":[217],"general":[219],"capabilities.":[220]},"counts_by_year":[{"year":2026,"cited_by_count":1}],"updated_date":"2026-08-08T01:25:22.217667","created_date":"2025-10-16T00:00:00"}