{"id":"https://openalex.org/W7204565985","doi":"https://doi.org/10.48550/arxiv.2608.27417","title":"Retrieval Heads Meet Vision: Uncovering How VLMs Locate and Extract Visual Information","display_name":"Retrieval Heads Meet Vision: Uncovering How VLMs Locate and Extract Visual Information","publication_year":2026,"publication_date":"2026-08-27","ids":{"openalex":"https://openalex.org/W7204565985","doi":"https://doi.org/10.48550/arxiv.2608.27417"},"language":"en","primary_location":{"id":"pmh:oai:arXiv.org:2608.27417","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2608.27417","pdf_url":"https://arxiv.org/pdf/2608.27417","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"},"volume":null,"type":"preprint","is_traditional":true,"indexed_in":["arxiv","datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://arxiv.org/pdf/2608.27417","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5136416265","display_name":"Chanho Park","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I157485424","display_name":"Korea Advanced Institute of Science and Technology","ror":"https://ror.org/05apxxy63","country_code":"KR","type":"education","lineage":["https://openalex.org/I157485424"]}],"countries":["KR"],"is_corresponding":false,"raw_author_name":"Park, Chanho","raw_affiliation_strings":["KAIST"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"KAIST","institution_ids":["https://openalex.org/I157485424"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5126220434","display_name":"Daehyeon Choi","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I157485424","display_name":"Korea Advanced Institute of Science and Technology","ror":"https://ror.org/05apxxy63","country_code":"KR","type":"education","lineage":["https://openalex.org/I157485424"]}],"countries":["KR"],"is_corresponding":false,"raw_author_name":"Choi, Daehyeon","raw_affiliation_strings":["KAIST"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"KAIST","institution_ids":["https://openalex.org/I157485424"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5143428824","display_name":"Jihyun Lee","orcid":"https://orcid.org/0009-0008-6684-1845","observed_orcids":["https://orcid.org/0009-0008-6684-1845"]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Lee, Jihyun","raw_affiliation_strings":["Independent Researcher"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Independent Researcher","institution_ids":[]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5004099860","display_name":"Minhyuk Sung","orcid":"https://orcid.org/0000-0001-7428-9570","observed_orcids":["https://orcid.org/0000-0001-7428-9570"]},"institutions":[{"id":"https://openalex.org/I157485424","display_name":"Korea Advanced Institute of Science and Technology","ror":"https://ror.org/05apxxy63","country_code":"KR","type":"education","lineage":["https://openalex.org/I157485424"]}],"countries":["KR"],"is_corresponding":false,"raw_author_name":"Sung, Minhyuk","raw_affiliation_strings":["KAIST"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"KAIST","institution_ids":["https://openalex.org/I157485424"]}]}],"institutions":[],"countries_distinct_count":1,"institutions_distinct_count":1,"corresponding_author_ids":[],"corresponding_institution_ids":[],"apc_list":null,"apc_paid":null,"fwci":null,"has_fulltext":true,"cited_by_count":3,"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":0.9994999766349792,"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":0.9994999766349792,"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/T11605","display_name":"Visual Attention and Saliency Detection","score":9.999999747378752e-05,"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/T10036","display_name":"Advanced Neural Network Applications","score":0.0,"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":0.968999981880188},{"id":"https://openalex.org/keywords/attention-heads","display_name":"attention heads","score":0.7110000252723694},{"id":"https://openalex.org/keywords/visual-grounding","display_name":"visual grounding","score":0.38600000739097595}],"concepts":[{"id":"https://openalex.org/C2777402240","wikidata":"https://www.wikidata.org/wiki/Q6783436","display_name":"Masking (illustration)","level":2,"score":0.7429999709129333},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.7358999848365784},{"id":"https://openalex.org/C2777096784","wikidata":"https://www.wikidata.org/wiki/Q3826351","display_name":"Referent","level":2,"score":0.6407999992370605},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.5753999948501587},{"id":"https://openalex.org/C26517878","wikidata":"https://www.wikidata.org/wiki/Q228039","display_name":"Key (lock)","level":2,"score":0.5090000033378601},{"id":"https://openalex.org/C204321447","wikidata":"https://www.wikidata.org/wiki/Q30642","display_name":"Natural language processing","level":1,"score":0.4383000135421753},{"id":"https://openalex.org/C23123220","wikidata":"https://www.wikidata.org/wiki/Q816826","display_name":"Information retrieval","level":1,"score":0.4228000044822693},{"id":"https://openalex.org/C115961682","wikidata":"https://www.wikidata.org/wiki/Q860623","display_name":"Image (mathematics)","level":2,"score":0.41519999504089355},{"id":"https://openalex.org/C1667742","wikidata":"https://www.wikidata.org/wiki/Q10927554","display_name":"Image retrieval","level":3,"score":0.40149998664855957},{"id":"https://openalex.org/C2779200073","wikidata":"https://www.wikidata.org/wiki/Q18395575","display_name":"Visual masking","level":4,"score":0.37529999017715454},{"id":"https://openalex.org/C2778572836","wikidata":"https://www.wikidata.org/wiki/Q380933","display_name":"Space (punctuation)","level":2,"score":0.3750999867916107},{"id":"https://openalex.org/C61797465","wikidata":"https://www.wikidata.org/wiki/Q1188986","display_name":"Term (time)","level":2,"score":0.3506999909877777},{"id":"https://openalex.org/C89611455","wikidata":"https://www.wikidata.org/wiki/Q6804646","display_name":"Mechanism (biology)","level":2,"score":0.34599998593330383},{"id":"https://openalex.org/C184337299","wikidata":"https://www.wikidata.org/wiki/Q1437428","display_name":"Semantics (computer science)","level":2,"score":0.32030001282691956},{"id":"https://openalex.org/C189391414","wikidata":"https://www.wikidata.org/wiki/Q7936579","display_name":"Visual Word","level":4,"score":0.31839999556541443},{"id":"https://openalex.org/C153180895","wikidata":"https://www.wikidata.org/wiki/Q7148389","display_name":"Pattern recognition (psychology)","level":2,"score":0.3174000084400177},{"id":"https://openalex.org/C2776401178","wikidata":"https://www.wikidata.org/wiki/Q12050496","display_name":"Feature (linguistics)","level":2,"score":0.31060001254081726},{"id":"https://openalex.org/C44291984","wikidata":"https://www.wikidata.org/wiki/Q1074173","display_name":"Question answering","level":2,"score":0.2937999963760376},{"id":"https://openalex.org/C31972630","wikidata":"https://www.wikidata.org/wiki/Q844240","display_name":"Computer vision","level":1,"score":0.2831999957561493},{"id":"https://openalex.org/C2776502983","wikidata":"https://www.wikidata.org/wiki/Q690182","display_name":"Contrast (vision)","level":2,"score":0.2766999900341034},{"id":"https://openalex.org/C207363949","wikidata":"https://www.wikidata.org/wiki/Q462915","display_name":"Visual space","level":3,"score":0.26980000734329224},{"id":"https://openalex.org/C178253425","wikidata":"https://www.wikidata.org/wiki/Q162668","display_name":"Visual perception","level":3,"score":0.2621000111103058},{"id":"https://openalex.org/C143271835","wikidata":"https://www.wikidata.org/wiki/Q254515","display_name":"Similitude","level":2,"score":0.2612999975681305},{"id":"https://openalex.org/C2777508537","wikidata":"https://www.wikidata.org/wiki/Q7936620","display_name":"Visual reasoning","level":2,"score":0.25450000166893005},{"id":"https://openalex.org/C36464697","wikidata":"https://www.wikidata.org/wiki/Q451553","display_name":"Visualization","level":2,"score":0.2542000114917755},{"id":"https://openalex.org/C184898388","wikidata":"https://www.wikidata.org/wiki/Q1435712","display_name":"Pairwise comparison","level":2,"score":0.25290000438690186},{"id":"https://openalex.org/C3004257","wikidata":"https://www.wikidata.org/wiki/Q17084606","display_name":"Correspondence problem","level":2,"score":0.25209999084472656}],"mesh":[],"locations_count":2,"locations":[{"id":"pmh:oai:arXiv.org:2608.27417","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2608.27417","pdf_url":"https://arxiv.org/pdf/2608.27417","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.2608.27417","doi":"https://doi.org/10.48550/arxiv.2608.27417","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2608.27417","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:2608.27417","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2608.27417","pdf_url":"https://arxiv.org/pdf/2608.27417","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":[],"funders":[],"has_content":{"pdf":true,"grobid_xml":true},"content_urls":{"pdf":"https://content.openalex.org/works/W7204565985.pdf","grobid_xml":"https://content.openalex.org/works/W7204565985.grobid-xml"},"referenced_works_count":11,"referenced_works":["https://openalex.org/W4365597196","https://openalex.org/W4387225514","https://openalex.org/W4405090824","https://openalex.org/W4405172832","https://openalex.org/W4405434026","https://openalex.org/W4407794264","https://openalex.org/W4416053382","https://openalex.org/W4416780398","https://openalex.org/W7077911270","https://openalex.org/W7128681975","https://openalex.org/W7133923255"],"related_works":["https://openalex.org/W6891918758","https://openalex.org/W7140080196","https://openalex.org/W7162504647","https://openalex.org/W4415254795","https://openalex.org/W7221542344","https://openalex.org/W4416036028","https://openalex.org/W4416359411","https://openalex.org/W7153978741","https://openalex.org/W4404988906","https://openalex.org/W7106597311"],"abstract_inverted_index":{"Vision-language":[0],"models":[1],"(VLMs)":[2],"can":[3],"locate":[4],"an":[5,46,216],"image":[6,78],"region":[7,118],"referred":[8],"to":[9,20,77,142],"by":[10,34,55,140],"a":[11,61,89,112],"text":[12,75,164],"prompt":[13],"and":[14,98,127,186,205,225],"route":[15],"the":[16,21,24,115,133,148,159],"corresponding":[17],"visual":[18,50,177],"evidence":[19],"output,":[22],"yet":[23],"internal":[25],"mechanism":[26,48],"behind":[27],"this":[28],"behavior":[29],"is":[30],"not":[31,171],"understood.":[32],"Inspired":[33],"retrieval":[35,165],"heads":[36,66,153],"in":[37,221],"large":[38],"language":[39],"models,":[40],"we":[41,83],"ask":[42],"whether":[43],"VLMs":[44,126,213],"contain":[45],"analogous":[47],"for":[49,73,163],"retrieval.":[51],"We":[52,101],"answer":[53],"affirmatively":[54],"introducing":[56],"Visual":[57],"Retrieval":[58],"Heads":[59],"(VRHs),":[60],"small":[62],"subset":[63],"of":[64,151],"attention":[65,106],"(about":[67],"1.7-2.6%)":[68],"that":[69,104,214],"are":[70,196,207],"causally":[71,211],"responsible":[72],"grounding":[74,138],"descriptions":[76],"regions.":[79],"To":[80],"find":[81],"them,":[82],"recast":[84],"existing":[85],"head-scoring":[86],"methods":[87],"under":[88],"unified":[90],"design":[91],"space":[92],"over":[93,114],"query":[94],"tokens,":[95],"key":[96],"aggregation,":[97],"cross-sample":[99],"aggregation.":[100],"then":[102],"show":[103],"scoring":[105],"from":[107],"output":[108,200],"prediction":[109],"tokens":[110],"with":[111],"sum":[113],"ground-truth":[116],"referent":[117],"most":[119],"reliably":[120],"identifies":[121],"causal":[122,181],"heads.":[123],"Across":[124],"eleven":[125],"five":[128],"referring-expression":[129],"benchmarks,":[130],"masking":[131,147],"only":[132],"top":[134],"20":[135],"VRHs":[136,167],"reduces":[137],"accuracy":[139],"up":[141],"80":[143],"percentage":[144],"points,":[145],"while":[146,202],"same":[149],"number":[150],"random":[152],"has":[154],"little":[155],"effect.":[156],"Beyond":[157],"replicating":[158],"causal-sparse-universal":[160],"triad":[161],"established":[162],"heads,":[166],"exhibit":[168],"several":[169],"properties":[170],"previously":[172],"reported:":[173],"they":[174,195,206],"generalize":[175],"across":[176,212],"reference":[178],"tasks,":[179],"remaining":[180],"on":[182],"attribute,":[183],"spatial,":[184],"counting,":[185],"visual-math":[187],"benchmarks":[188],"despite":[189],"being":[190],"discovered":[191],"through":[192],"bounding-box":[193],"prediction;":[194],"functionally":[197],"specific,":[198],"preserving":[199],"format":[201],"corrupting":[203],"localization;":[204],"architecturally":[208],"shared,":[209],"transferring":[210],"share":[215],"LLM":[217],"backbone":[218],"but":[219],"differ":[220],"vision":[222],"encoder,":[223],"projector,":[224],"instruction":[226],"tuning.":[227]},"counts_by_year":[{"year":2026,"cited_by_count":3}],"updated_date":"2026-11-07T11:25:07.637180","created_date":"2026-08-29T00:00:00"}