{"id":"https://openalex.org/W7160053354","doi":"https://doi.org/10.1109/iccv51701.2025.02094","title":"TAB: Transformer Attention Bottlenecks Enable User Intervention and Debugging in Vision-Language Models","display_name":"TAB: Transformer Attention Bottlenecks Enable User Intervention and Debugging in Vision-Language Models","publication_year":2025,"publication_date":"2025-10-19","ids":{"openalex":"https://openalex.org/W7160053354","doi":"https://doi.org/10.1109/iccv51701.2025.02094"},"language":"en","primary_location":{"id":"doi:10.1109/iccv51701.2025.02094","doi":"https://doi.org/10.1109/iccv51701.2025.02094","is_oa":false,"landing_page_url":"https://doi.org/10.1109/iccv51701.2025.02094","pdf_url":null,"source":{"id":"https://openalex.org/S4363607764","display_name":"IEEE/CVF International Conference on Computer Vision (ICCV)","issn_l":null,"issn":null,"is_oa":false,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":null,"host_organization_name":null,"host_organization_lineage":[],"host_organization_lineage_names":[],"type":"conference"},"license":null,"license_id":null,"version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"2025 IEEE/CVF International Conference on Computer Vision (ICCV)","raw_type":"proceedings-article"},"volume":null,"type":"conference-paper","indexed_in":["arxiv","crossref","datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://arxiv.org/pdf/2412.18675","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5103305861","display_name":"Pooyan Rahmanzadehgervi","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I82497590","display_name":"Auburn University","ror":"https://ror.org/02v80fc35","country_code":"US","type":"education","lineage":["https://openalex.org/I82497590"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Pooyan Rahmanzadehgervi","raw_affiliation_strings":["Auburn University"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Auburn University","institution_ids":["https://openalex.org/I82497590"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5101936580","display_name":"Hung Huy Nguyen","orcid":"https://orcid.org/0000-0003-3217-2723","observed_orcids":["https://orcid.org/0000-0003-3217-2723"]},"institutions":[{"id":"https://openalex.org/I82497590","display_name":"Auburn University","ror":"https://ror.org/02v80fc35","country_code":"US","type":"education","lineage":["https://openalex.org/I82497590"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Hung Huy Nguyen","raw_affiliation_strings":["Auburn University"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Auburn University","institution_ids":["https://openalex.org/I82497590"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5135111221","display_name":"Rosanne Liu","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I4210090411","display_name":"Google DeepMind (United Kingdom)","ror":"https://ror.org/00971b260","country_code":"GB","type":"company","lineage":["https://openalex.org/I4210090411","https://openalex.org/I4210128969"]}],"countries":["GB"],"is_corresponding":false,"raw_author_name":"Rosanne Liu","raw_affiliation_strings":["Google DeepMind, ML Collective"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Google DeepMind, ML Collective","institution_ids":["https://openalex.org/I4210090411"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5022192060","display_name":"Long Mai","orcid":"https://orcid.org/0009-0001-1403-9884","observed_orcids":["https://orcid.org/0009-0001-1403-9884"]},"institutions":[{"id":"https://openalex.org/I1306409833","display_name":"Adobe Systems (United States)","ror":"https://ror.org/059tvcg64","country_code":"US","type":"company","lineage":["https://openalex.org/I1306409833"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Long Mai","raw_affiliation_strings":["Adobe Research"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Adobe Research","institution_ids":["https://openalex.org/I1306409833"]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5113284752","display_name":"Anh Totti Nguyen","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I82497590","display_name":"Auburn University","ror":"https://ror.org/02v80fc35","country_code":"US","type":"education","lineage":["https://openalex.org/I82497590"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Anh Totti Nguyen","raw_affiliation_strings":["Auburn University"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Auburn University","institution_ids":["https://openalex.org/I82497590"]}]}],"institutions":[],"countries_distinct_count":2,"institutions_distinct_count":3,"corresponding_author_ids":[],"corresponding_institution_ids":[],"apc_list":null,"apc_paid":null,"fwci":0.7464,"has_fulltext":false,"cited_by_count":1,"citation_normalized_percentile":{"value":0.80846113,"is_in_top_1_percent":false,"is_in_top_10_percent":false},"cited_by_percentile_year":{"min":90,"max":94},"biblio":{"volume":null,"issue":null,"first_page":"22551","last_page":"22562"},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T11714","display_name":"Multimodal Machine Learning Applications","score":0.9907000064849854,"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"}},"topics":[{"id":"https://openalex.org/T11714","display_name":"Multimodal Machine Learning Applications","score":0.9907000064849854,"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/T12026","display_name":"Explainable Artificial Intelligence (XAI)","score":0.004000000189989805,"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/T10036","display_name":"Advanced Neural Network Applications","score":0.0012000000569969416,"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"}}],"keywords":[{"id":"https://openalex.org/keywords/vision-language-models","display_name":"vision-language models","score":0.8349999785423279},{"id":"https://openalex.org/keywords/user-intervention","display_name":"user intervention","score":0.902999997138977},{"id":"https://openalex.org/keywords/debugging","display_name":"debugging","score":0.8360000252723694},{"id":"https://openalex.org/keywords/transformer-attention","display_name":"transformer attention","score":0.40299999713897705},{"id":"https://openalex.org/keywords/interpretability","display_name":"interpretability","score":0.5830000042915344}],"concepts":[{"id":"https://openalex.org/C168065819","wikidata":"https://www.wikidata.org/wiki/Q845566","display_name":"Debugging","level":2,"score":0.7056000232696533},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.6556000113487244},{"id":"https://openalex.org/C107457646","wikidata":"https://www.wikidata.org/wiki/Q207434","display_name":"Human\u2013computer interaction","level":1,"score":0.398499995470047},{"id":"https://openalex.org/C89505385","wikidata":"https://www.wikidata.org/wiki/Q47146","display_name":"User interface","level":2,"score":0.390500009059906},{"id":"https://openalex.org/C115903868","wikidata":"https://www.wikidata.org/wiki/Q80993","display_name":"Software engineering","level":1,"score":0.35190001130104065},{"id":"https://openalex.org/C66322947","wikidata":"https://www.wikidata.org/wiki/Q11658","display_name":"Transformer","level":3,"score":0.3255999982357025},{"id":"https://openalex.org/C149635348","wikidata":"https://www.wikidata.org/wiki/Q193040","display_name":"Embedded system","level":1,"score":0.2865999937057495},{"id":"https://openalex.org/C26517878","wikidata":"https://www.wikidata.org/wiki/Q228039","display_name":"Key (lock)","level":2,"score":0.2818000018596649},{"id":"https://openalex.org/C127413603","wikidata":"https://www.wikidata.org/wiki/Q11023","display_name":"Engineering","level":0,"score":0.27810001373291016},{"id":"https://openalex.org/C67712803","wikidata":"https://www.wikidata.org/wiki/Q7901853","display_name":"User modeling","level":3,"score":0.2736999988555908},{"id":"https://openalex.org/C186967261","wikidata":"https://www.wikidata.org/wiki/Q5082128","display_name":"Mobile device","level":2,"score":0.2700999975204468},{"id":"https://openalex.org/C2780665704","wikidata":"https://www.wikidata.org/wiki/Q959298","display_name":"Intervention (counseling)","level":2,"score":0.25949999690055847},{"id":"https://openalex.org/C98045186","wikidata":"https://www.wikidata.org/wiki/Q205663","display_name":"Process (computing)","level":2,"score":0.25600001215934753}],"mesh":[],"locations_count":3,"locations":[{"id":"doi:10.1109/iccv51701.2025.02094","doi":"https://doi.org/10.1109/iccv51701.2025.02094","is_oa":false,"landing_page_url":"https://doi.org/10.1109/iccv51701.2025.02094","pdf_url":null,"source":{"id":"https://openalex.org/S4363607764","display_name":"IEEE/CVF International Conference on Computer Vision (ICCV)","issn_l":null,"issn":null,"is_oa":false,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":null,"host_organization_name":null,"host_organization_lineage":[],"host_organization_lineage_names":[],"type":"conference"},"license":null,"license_id":null,"version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"2025 IEEE/CVF International Conference on Computer Vision (ICCV)","raw_type":"proceedings-article"},{"id":"pmh:oai:arXiv.org:2412.18675","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2412.18675","pdf_url":"https://arxiv.org/pdf/2412.18675","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.2412.18675","doi":"https://doi.org/10.48550/arxiv.2412.18675","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2412.18675","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:2412.18675","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2412.18675","pdf_url":"https://arxiv.org/pdf/2412.18675","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":[{"id":"https://openalex.org/G3464416950","display_name":"CRII: RI: Testing and Interpreting Image-based Computer Vision Models in 3D Space","funder_award_id":"1850117","funder_id":"https://openalex.org/F4320306076","funder_display_name":"National Science Foundation"}],"funders":[{"id":"https://openalex.org/F4320306076","display_name":"National Science Foundation","ror":"https://ror.org/021nxhr62"}],"has_content":{"grobid_xml":false,"pdf":false},"content_urls":null,"referenced_works_count":49,"referenced_works":["https://openalex.org/W1956340063","https://openalex.org/W2101105183","https://openalex.org/W2142996775","https://openalex.org/W2221625691","https://openalex.org/W2295107390","https://openalex.org/W2606462007","https://openalex.org/W2626778328","https://openalex.org/W2776207810","https://openalex.org/W2889017949","https://openalex.org/W2938232066","https://openalex.org/W2949554687","https://openalex.org/W3004725381","https://openalex.org/W3022265721","https://openalex.org/W3035561175","https://openalex.org/W3046090006","https://openalex.org/W3082665562","https://openalex.org/W3092088514","https://openalex.org/W3128528696","https://openalex.org/W3132191748","https://openalex.org/W3140300848","https://openalex.org/W3151130473","https://openalex.org/W3159344844","https://openalex.org/W3159481202","https://openalex.org/W3172942063","https://openalex.org/W3173550280","https://openalex.org/W3199003182","https://openalex.org/W4225739172","https://openalex.org/W4285192809","https://openalex.org/W4288083516","https://openalex.org/W4289537913","https://openalex.org/W4312770831","https://openalex.org/W4319300177","https://openalex.org/W4385245495","https://openalex.org/W4386071596","https://openalex.org/W4386076127","https://openalex.org/W4390190626","https://openalex.org/W4393150383","https://openalex.org/W4394597936","https://openalex.org/W4401042250","https://openalex.org/W4402716012","https://openalex.org/W4402727764","https://openalex.org/W4402754023","https://openalex.org/W4403089452","https://openalex.org/W4404784092","https://openalex.org/W4406844129","https://openalex.org/W4409262175","https://openalex.org/W4413144238","https://openalex.org/W7133192941","https://openalex.org/W7133236469"],"related_works":["https://openalex.org/W4293421034","https://openalex.org/W3153758184","https://openalex.org/W4312791030","https://openalex.org/W2934032625","https://openalex.org/W6967207777","https://openalex.org/W7140080196","https://openalex.org/W2939556020","https://openalex.org/W7101415090","https://openalex.org/W4404931900","https://openalex.org/W4360995299"],"abstract_inverted_index":{"Multi-head":[0],"self-attention":[1],"(MHSA)":[2],"is":[3,92,97,145,158],"a":[4,9,43,47,111],"key":[5],"component":[6],"of":[7,35,42,119],"Transformers,":[8],"widely":[10],"popular":[11],"architecture":[12,161],"in":[13,140,147,151],"both":[14],"language":[15],"and":[16,69,103,150],"vision.":[17],"Multiple":[18],"heads":[19],"intuitively":[20],"enable":[21,163],"different":[22],"parallel":[23],"processes":[24],"over":[25,79],"the":[26,33,40,57,76,89,101,104,117,143,159],"same":[27],"input.":[28],"Yet,":[29],"they":[30],"also":[31],"obscure":[32],"attribution":[34],"each":[36],"input":[37],"patch":[38],"to":[39,61,82,110,126,137,162,165],"output":[41],"model.":[44],"We":[45],"propose":[46],"novel":[48],"1-head":[49],"Transformer":[50],"Attention":[51],"Bottleneck":[52],"(TAB)":[53],"layer,":[54],"inserted":[55],"after":[56],"traditional":[58],"MHSA":[59],"architecture,":[60],"serve":[62],"as":[63],"an":[64],"attention":[65,78,91],"bottleneck":[66,144],"for":[67],"interpretability":[68],"intervention.":[70],"Unlike":[71],"standard":[72],"self-attention,":[73],"TAB":[74,125,157],"constrains":[75],"total":[77,90],"all":[80],"patches":[81],"$\\in":[83],"[0,":[84],"1]$.":[85],"That":[86],"is,":[87],"when":[88,153],"0,":[93],"no":[94,154],"visual":[95],"information":[96],"propagated":[98],"further":[99],"into":[100],"network,":[102],"vision-language":[105],"model":[106],"(VLM)":[107],"would":[108],"default":[109],"generic,":[112],"image-independent":[113],"response.":[114],"To":[115],"demonstrate":[116],"advantages":[118],"TAB,":[120],"we":[121],"train":[122],"VLMs":[123,139],"with":[124],"perform":[127,135],"image-difference":[128],"captioning.":[129],"Over":[130],"three":[131],"datasets,":[132],"our":[133],"models":[134],"similarly":[136],"baseline":[138],"captioning":[141],"but":[142],"superior":[146],"localizing":[148],"changes":[149,155],"identifying":[152],"occur.":[156],"first":[160],"users":[164],"debug":[166],"by":[167,175],"editing":[168],"attention,":[169],"which":[170],"often":[171],"produces":[172],"expected":[173],"outputs":[174],"VLMs.":[176]},"counts_by_year":[{"year":2025,"cited_by_count":1}],"updated_date":"2026-10-29T09:56:32.521200","created_date":"2026-05-05T00:00:00"}