{"id":"https://openalex.org/W7163640489","doi":"https://doi.org/10.48550/arxiv.2606.05843","title":"Mechanistic Insights into Functional Sparsity in Multimodal LLMs via CoRe Heads","display_name":"Mechanistic Insights into Functional Sparsity in Multimodal LLMs via CoRe Heads","publication_year":2026,"publication_date":"2026-06-04","ids":{"openalex":"https://openalex.org/W7163640489","doi":"https://doi.org/10.48550/arxiv.2606.05843"},"language":null,"primary_location":{"id":"doi:10.48550/arxiv.2606.05843","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2606.05843","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.2606.05843","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5137961643","display_name":"Ruoxi Sun","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Sun, Ruoxi","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5123981501","display_name":"Quantong Qiu","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Qiu, Quantong","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5138007758","display_name":"Juntao Li","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Li, Juntao","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5137952630","display_name":"Zecheng Tang","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Tang, Zecheng","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5063901345","display_name":"Yihang Lou","orcid":"https://orcid.org/0000-0002-8143-389X"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Lou, Yihang","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5137954785","display_name":"Min Zhang","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Zhang, Min","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":0.9728000164031982,"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":0.9728000164031982,"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":0.003599999938160181,"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/T11307","display_name":"Domain Adaptation and Few-Shot Learning","score":0.0027000000700354576,"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/interpretability","display_name":"Interpretability","score":0.958899974822998},{"id":"https://openalex.org/keywords/inference","display_name":"Inference","score":0.5443999767303467},{"id":"https://openalex.org/keywords/core","display_name":"Core (optical fiber)","score":0.5113999843597412},{"id":"https://openalex.org/keywords/task","display_name":"Task (project management)","score":0.5080999732017517},{"id":"https://openalex.org/keywords/metric","display_name":"Metric (unit)","score":0.5005000233650208},{"id":"https://openalex.org/keywords/property","display_name":"Property (philosophy)","score":0.4180000126361847},{"id":"https://openalex.org/keywords/similarity","display_name":"Similarity (geometry)","score":0.3449999988079071}],"concepts":[{"id":"https://openalex.org/C2781067378","wikidata":"https://www.wikidata.org/wiki/Q17027399","display_name":"Interpretability","level":2,"score":0.958899974822998},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.6970999836921692},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.6182000041007996},{"id":"https://openalex.org/C2776214188","wikidata":"https://www.wikidata.org/wiki/Q408386","display_name":"Inference","level":2,"score":0.5443999767303467},{"id":"https://openalex.org/C2164484","wikidata":"https://www.wikidata.org/wiki/Q5170150","display_name":"Core (optical fiber)","level":2,"score":0.5113999843597412},{"id":"https://openalex.org/C2780451532","wikidata":"https://www.wikidata.org/wiki/Q759676","display_name":"Task (project management)","level":2,"score":0.5080999732017517},{"id":"https://openalex.org/C176217482","wikidata":"https://www.wikidata.org/wiki/Q860554","display_name":"Metric (unit)","level":2,"score":0.5005000233650208},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.446399986743927},{"id":"https://openalex.org/C189950617","wikidata":"https://www.wikidata.org/wiki/Q937228","display_name":"Property (philosophy)","level":2,"score":0.4180000126361847},{"id":"https://openalex.org/C103278499","wikidata":"https://www.wikidata.org/wiki/Q254465","display_name":"Similarity (geometry)","level":3,"score":0.3449999988079071},{"id":"https://openalex.org/C63479239","wikidata":"https://www.wikidata.org/wiki/Q7353546","display_name":"Robustness (evolution)","level":3,"score":0.3237000107765198},{"id":"https://openalex.org/C175154964","wikidata":"https://www.wikidata.org/wiki/Q380077","display_name":"Task analysis","level":3,"score":0.3084000051021576},{"id":"https://openalex.org/C66024118","wikidata":"https://www.wikidata.org/wiki/Q1122506","display_name":"Computational model","level":2,"score":0.3050999939441681},{"id":"https://openalex.org/C203005215","wikidata":"https://www.wikidata.org/wiki/Q79798","display_name":"Machine translation","level":2,"score":0.3012000024318695},{"id":"https://openalex.org/C204321447","wikidata":"https://www.wikidata.org/wiki/Q30642","display_name":"Natural language processing","level":1,"score":0.28999999165534973},{"id":"https://openalex.org/C158600405","wikidata":"https://www.wikidata.org/wiki/Q5054566","display_name":"Causal inference","level":2,"score":0.28349998593330383},{"id":"https://openalex.org/C36464697","wikidata":"https://www.wikidata.org/wiki/Q451553","display_name":"Visualization","level":2,"score":0.2824000120162964},{"id":"https://openalex.org/C2987834672","wikidata":"https://www.wikidata.org/wiki/Q4677630","display_name":"Action recognition","level":3,"score":0.26930001378059387},{"id":"https://openalex.org/C123657996","wikidata":"https://www.wikidata.org/wiki/Q12271","display_name":"Architecture","level":2,"score":0.2540000081062317}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.48550/arxiv.2606.05843","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2606.05843","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.2606.05843","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2606.05843","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":"Quality Education","id":"https://metadata.un.org/sdg/4","score":0.4943186938762665}],"awards":[],"funders":[],"has_content":{"grobid_xml":false,"pdf":false},"content_urls":null,"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"While":[0],"Multimodal":[1],"Large":[2],"Language":[3],"Models":[4],"(MLLMs)":[5],"demonstrate":[6,110],"remarkable":[7],"proficiency":[8],"on":[9],"complex":[10],"vision-language":[11],"tasks,":[12],"the":[13,111,119,143,174],"mechanisms":[14],"by":[15],"which":[16],"they":[17],"extract":[18],"query-relevant":[19],"visual":[20,79],"features":[21],"from":[22],"complex,":[23],"noisy":[24],"contexts":[25],"remain":[26],"opaque.":[27],"In":[28],"this":[29,151],"paper,":[30],"we":[31,59,84],"present":[32],"an":[33],"in-depth":[34],"interpretability":[35,179],"study":[36],"that":[37,149,185],"uncovers":[38],"a":[39,51,63,86,165,182],"profound":[40],"structural":[41,166],"property":[42],"within":[43,171],"MLLMs:":[44],"functional":[45,88,169],"sparsity":[46,153,170],"in":[47,128],"cross-modal":[48],"retrieval.":[49],"Leveraging":[50],"token-level":[52],"metric":[53],"termed":[54],"Retrieval":[55,74],"Attention":[56],"Mass":[57],"(RAM),":[58],"identify":[60],"and":[61,81,180,191],"characterize":[62],"highly":[64],"specialized":[65,115],"subset":[66],"of":[67,113,122,145,168,177],"attention":[68,102],"heads,":[69,147],"referred":[70],"to":[71],"as":[72,93],"Context-aware":[73],"(CoRe)":[75],"heads.":[76,116],"Across":[77],"diverse":[78],"domains":[80],"model":[82,192],"scales,":[83],"observe":[85],"clear":[87],"division:":[89],"CoRe":[90,123,146],"heads":[91,100,124,135],"act":[92],"dedicated":[94],"information":[95],"extractors,":[96],"while":[97,157],"most":[98],"other":[99],"distribute":[101],"over":[103],"broader":[104],"contextual":[105],"regions.":[106],"Causal":[107],"interventions":[108],"further":[109],"necessity":[112],"these":[114],"Ablating":[117],"only":[118],"top":[120],"5%":[121],"causes":[125],"significant":[126],"degradation":[127],"multimodal":[129],"reasoning":[130],"performance,":[131],"whereas":[132],"ablating":[133],"lower-ranked":[134],"has":[136],"minimal":[137],"effect.":[138],"Moreover,":[139],"acceleration":[140],"experiments":[141],"validate":[142],"utility":[144],"showing":[148],"leveraging":[150],"localized":[152],"significantly":[154],"accelerates":[155],"inference":[156],"maintaining":[158],"robust":[159],"task":[160],"performance.":[161],"Our":[162],"findings":[163],"reveal":[164],"principle":[167],"MLLMs,":[172],"refining":[173],"current":[175],"understanding":[176],"mechanistic":[178],"laying":[181],"theoretical":[183],"foundation":[184],"can":[186],"inspire":[187],"future":[188],"architecture":[189],"design":[190],"optimization.":[193]},"counts_by_year":[],"updated_date":"2026-07-01T06:00:48.157686","created_date":"2026-06-06T00:00:00"}