{"id":"https://openalex.org/W7225273544","doi":"https://doi.org/10.48550/arxiv.2611.04903","title":"Heads That Write, Not Just Point: Image Retrieval Heads in Vision-Language Models","display_name":"Heads That Write, Not Just Point: Image Retrieval Heads in Vision-Language Models","publication_year":2026,"publication_date":"2026-11-03","ids":{"openalex":"https://openalex.org/W7225273544","doi":"https://doi.org/10.48550/arxiv.2611.04903"},"language":"en","primary_location":{"id":"doi:10.48550/arxiv.2611.04903","doi":"https://doi.org/10.48550/arxiv.2611.04903","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2611.04903","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":null,"license_id":null,"version":null,"is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"Preprint"},"volume":null,"type":"preprint","indexed_in":["datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://doi.org/10.48550/arxiv.2611.04903","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5136638517","display_name":"Junsung Park","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Park, Junsung","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5160633162","display_name":"Sangwon Yu","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Yu, Sangwon","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5160594463","display_name":"Donghun Kang","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Kang, Donghun","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5033377277","display_name":"Yeongtak Oh","orcid":"https://orcid.org/0000-0002-3642-1830","observed_orcids":["https://orcid.org/0000-0002-3642-1830"]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Oh, Yeongtak","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5135149561","display_name":"Jaewon Jeong","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Jeong, Jaewon","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5137824139","display_name":"Gyeongtae Yoo","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Yoo, Gyeongtae","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5051378764","display_name":"Han Cheol Moon","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Moon, Han Cheol","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5160581236","display_name":"Jungbeom Lee","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Lee, Jungbeom","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5160692699","display_name":"Sungroh Yoon","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Yoon, Sungroh","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/T11714","display_name":"Multimodal Machine Learning Applications","score":0.9984999895095825,"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":0.9984999895095825,"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/T10028","display_name":"Topic Modeling","score":0.0003000000142492354,"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":0.00019999999494757503,"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":[],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.7670000195503235},{"id":"https://openalex.org/C150202949","wikidata":"https://www.wikidata.org/wiki/Q107602","display_name":"Pointer (user interface)","level":2,"score":0.7075999975204468},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.6194000244140625},{"id":"https://openalex.org/C31972630","wikidata":"https://www.wikidata.org/wiki/Q844240","display_name":"Computer vision","level":1,"score":0.6035000085830688},{"id":"https://openalex.org/C1667742","wikidata":"https://www.wikidata.org/wiki/Q10927554","display_name":"Image retrieval","level":3,"score":0.554099977016449},{"id":"https://openalex.org/C2776214188","wikidata":"https://www.wikidata.org/wiki/Q408386","display_name":"Inference","level":2,"score":0.42250001430511475},{"id":"https://openalex.org/C115961682","wikidata":"https://www.wikidata.org/wiki/Q860623","display_name":"Image (mathematics)","level":2,"score":0.3885999917984009},{"id":"https://openalex.org/C28719098","wikidata":"https://www.wikidata.org/wiki/Q44946","display_name":"Point (geometry)","level":2,"score":0.37790000438690186},{"id":"https://openalex.org/C175291020","wikidata":"https://www.wikidata.org/wiki/Q1156822","display_name":"Offset (computer science)","level":2,"score":0.3727000057697296},{"id":"https://openalex.org/C23123220","wikidata":"https://www.wikidata.org/wiki/Q816826","display_name":"Information retrieval","level":1,"score":0.31790000200271606},{"id":"https://openalex.org/C153180895","wikidata":"https://www.wikidata.org/wiki/Q7148389","display_name":"Pattern recognition (psychology)","level":2,"score":0.3165000081062317},{"id":"https://openalex.org/C98045186","wikidata":"https://www.wikidata.org/wiki/Q205663","display_name":"Process (computing)","level":2,"score":0.31049999594688416},{"id":"https://openalex.org/C97256817","wikidata":"https://www.wikidata.org/wiki/Q1462316","display_name":"Spurious relationship","level":2,"score":0.30649998784065247},{"id":"https://openalex.org/C200873422","wikidata":"https://www.wikidata.org/wiki/Q5448821","display_name":"Filling-in","level":2,"score":0.29649999737739563},{"id":"https://openalex.org/C9417928","wikidata":"https://www.wikidata.org/wiki/Q1070689","display_name":"Image processing","level":3,"score":0.29249998927116394},{"id":"https://openalex.org/C2780312720","wikidata":"https://www.wikidata.org/wiki/Q5689100","display_name":"Head (geology)","level":2,"score":0.29190000891685486},{"id":"https://openalex.org/C2776760102","wikidata":"https://www.wikidata.org/wiki/Q5139990","display_name":"Code (set theory)","level":3,"score":0.25609999895095825},{"id":"https://openalex.org/C184898388","wikidata":"https://www.wikidata.org/wiki/Q1435712","display_name":"Pairwise comparison","level":2,"score":0.2513999938964844}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.48550/arxiv.2611.04903","doi":"https://doi.org/10.48550/arxiv.2611.04903","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2611.04903","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":null,"license_id":null,"version":null,"is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"Preprint"}],"best_oa_location":{"id":"doi:10.48550/arxiv.2611.04903","doi":"https://doi.org/10.48550/arxiv.2611.04903","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2611.04903","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":null,"license_id":null,"version":null,"is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"Preprint"},"sustainable_development_goals":[],"study_designs":[],"awards":[],"funders":[],"has_content":{"pdf":false,"grobid_xml":false},"content_urls":null,"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Large":[0],"vision-language":[1,85],"models":[2],"(LVLMs)":[3],"can":[4],"process":[5],"interleaved":[6],"text":[7],"and":[8,38,141,161,230,235],"multiple":[9],"images":[10],"in":[11,198],"a":[12,24,41,77,88,97,108,124,133],"single":[13,78],"context,":[14],"but":[15],"how":[16],"they":[17],"internally":[18],"identify":[19,60,145,154],"the":[20,68,75,155,159,177,188,227],"image":[21,36,48,100,135],"relevant":[22],"to":[23,67,118,144,153,172],"question":[25],"remains":[26],"poorly":[27],"understood.":[28],"We":[29],"study":[30],"this":[31,199],"grounding":[32],"step":[33],"as":[34],"in-context":[35],"retrieval":[37,91,101],"introduce":[39,123],"ICIR-MCQ,":[40],"controlled":[42],"probe":[43],"for":[44,192],"head-level":[45],"analysis":[46],"of":[47],"retrieval.":[49],"Using":[50],"each":[51,129,181],"head's":[52],"attention":[53],"mass":[54],"over":[55],"image-token":[56],"spans,":[57],"we":[58,122,131],"first":[59],"sparse":[61],"pointer":[62,79,160,229],"heads":[63,72,174,179,189,203,223],"that":[64,94,187],"reliably":[65],"point":[66],"target":[69,156],"image.":[70,157],"These":[71],"transfer":[73],"beyond":[74],"probe:":[76],"head":[80,109,167],"is":[81],"competitive":[82],"with":[83,169],"external":[84],"retrievers":[86],"on":[87,117,137],"naturalistic":[89],"multi-image":[90,194],"benchmark,":[92],"showing":[93],"LVLMs":[95],"contain":[96,150],"strong":[98],"internal":[99],"signal.":[102],"Attention":[103],"alone,":[104],"however,":[105],"shows":[106],"where":[107],"reads":[110],"from,":[111],"not":[112],"what":[113],"its":[114,138],"output":[115,140],"passes":[116],"later":[119],"layers,":[120],"so":[121],"complementary":[125],"output-side":[126,231],"analysis.":[127],"For":[128],"head,":[130],"train":[132],"lightweight":[134],"retriever":[136],"value-weighted":[139],"use":[142],"it":[143],"writer":[146,162,232],"heads,":[147],"whose":[148],"outputs":[149],"enough":[151],"information":[152],"Surprisingly,":[158],"criteria":[163],"select":[164],"substantially":[165],"different":[166],"sets,":[168],"only":[170,206],"35":[171],"47":[173],"overlapping":[175],"among":[176],"top-100":[178],"under":[180],"criterion.":[182],"Targeted":[183],"set-partition":[184],"ablations":[185],"show":[186,209],"most":[190,217],"important":[191],"standard":[193],"inference":[195],"are":[196],"concentrated":[197],"pointer-writer":[200],"intersection,":[201],"while":[202],"selected":[204,224],"by":[205,225],"one":[207],"criterion":[208],"less":[210],"consistent":[211,218],"separation":[212],"from":[213,222],"random-head":[214],"controls.":[215],"The":[216],"causal":[219],"effects":[220],"arise":[221],"both":[226],"attention-based":[228],"criteria.":[233],"Code":[234],"data:":[236],"https://github.com/parkquasar/Image-Retrieval-Heads.":[237]},"counts_by_year":[],"updated_date":"2026-11-05T12:25:35.769452","created_date":"2026-11-05T00:00:00"}