{"id":"https://openalex.org/W4415345266","doi":"https://doi.org/10.48550/arxiv.2507.03865","title":"OrthoRank: Token Selection via Sink Token Orthogonality for Efficient LLM inference","display_name":"OrthoRank: Token Selection via Sink Token Orthogonality for Efficient LLM inference","publication_year":2025,"publication_date":"2025-07-05","ids":{"openalex":"https://openalex.org/W4415345266","doi":"https://doi.org/10.48550/arxiv.2507.03865"},"language":"en","primary_location":{"id":"pmh:oai:arXiv.org:2507.03865","is_oa":true,"landing_page_url":"http://arxiv.org/abs/2507.03865","pdf_url":"https://arxiv.org/pdf/2507.03865","source":{"id":"https://openalex.org/S4393918464","display_name":"ArXiv.org","issn_l":"2331-8422","issn":["2331-8422"],"is_oa":true,"is_in_doaj":false,"is_core":false,"host_organization":null,"host_organization_name":null,"host_organization_lineage":[],"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"},"type":"preprint","indexed_in":["arxiv","datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://arxiv.org/pdf/2507.03865","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":null,"display_name":"Shin, Seungjun","orcid":null},"institutions":[],"countries":[],"is_corresponding":true,"raw_author_name":"Shin, Seungjun","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5101177031","display_name":"Jaehoon Oh","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Oh, Jaehoon","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5059766169","display_name":"Do\u2010Youn Oh","orcid":"https://orcid.org/0000-0003-1663-9901"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Oh, Dokwan","raw_affiliation_strings":[],"affiliations":[]}],"institutions":[],"countries_distinct_count":0,"institutions_distinct_count":3,"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/T10181","display_name":"Natural Language Processing Techniques","score":0.9970999956130981,"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"}},"topics":[{"id":"https://openalex.org/T10181","display_name":"Natural Language Processing Techniques","score":0.9970999956130981,"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/T10201","display_name":"Speech Recognition and Synthesis","score":0.9516000151634216,"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/T10601","display_name":"Handwritten Text Recognition Techniques","score":0.9513999819755554,"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/security-token","display_name":"Security token","score":0.8564000129699707},{"id":"https://openalex.org/keywords/sink","display_name":"Sink (geography)","score":0.578499972820282},{"id":"https://openalex.org/keywords/token-passing","display_name":"Token passing","score":0.5407999753952026},{"id":"https://openalex.org/keywords/inference","display_name":"Inference","score":0.41690000891685486},{"id":"https://openalex.org/keywords/equalizer","display_name":"Equalizer","score":0.34439998865127563}],"concepts":[{"id":"https://openalex.org/C48145219","wikidata":"https://www.wikidata.org/wiki/Q1335365","display_name":"Security token","level":2,"score":0.8564000129699707},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.7084000110626221},{"id":"https://openalex.org/C143050476","wikidata":"https://www.wikidata.org/wiki/Q194502","display_name":"Sink (geography)","level":2,"score":0.578499972820282},{"id":"https://openalex.org/C115067241","wikidata":"https://www.wikidata.org/wiki/Q1639854","display_name":"Token passing","level":3,"score":0.5407999753952026},{"id":"https://openalex.org/C2776214188","wikidata":"https://www.wikidata.org/wiki/Q408386","display_name":"Inference","level":2,"score":0.41690000891685486},{"id":"https://openalex.org/C67545415","wikidata":"https://www.wikidata.org/wiki/Q5384218","display_name":"Equalizer","level":3,"score":0.34439998865127563},{"id":"https://openalex.org/C18664526","wikidata":"https://www.wikidata.org/wiki/Q7651187","display_name":"Suzuki-Kasami algorithm","level":3,"score":0.3334999978542328},{"id":"https://openalex.org/C102462485","wikidata":"https://www.wikidata.org/wiki/Q904595","display_name":"Token bus network","level":3,"score":0.30640000104904175},{"id":"https://openalex.org/C31258907","wikidata":"https://www.wikidata.org/wiki/Q1301371","display_name":"Computer network","level":1,"score":0.3021000027656555},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.2906000018119812},{"id":"https://openalex.org/C11413529","wikidata":"https://www.wikidata.org/wiki/Q8366","display_name":"Algorithm","level":1,"score":0.2671999931335449}],"mesh":[],"locations_count":2,"locations":[{"id":"pmh:oai:arXiv.org:2507.03865","is_oa":true,"landing_page_url":"http://arxiv.org/abs/2507.03865","pdf_url":"https://arxiv.org/pdf/2507.03865","source":{"id":"https://openalex.org/S4393918464","display_name":"ArXiv.org","issn_l":"2331-8422","issn":["2331-8422"],"is_oa":true,"is_in_doaj":false,"is_core":false,"host_organization":null,"host_organization_name":null,"host_organization_lineage":[],"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.2507.03865","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2507.03865","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":null,"license_id":null,"version":null,"is_accepted":false,"is_published":null,"raw_source_name":null,"raw_type":"article"}],"best_oa_location":{"id":"pmh:oai:arXiv.org:2507.03865","is_oa":true,"landing_page_url":"http://arxiv.org/abs/2507.03865","pdf_url":"https://arxiv.org/pdf/2507.03865","source":{"id":"https://openalex.org/S4393918464","display_name":"ArXiv.org","issn_l":"2331-8422","issn":["2331-8422"],"is_oa":true,"is_in_doaj":false,"is_core":false,"host_organization":null,"host_organization_name":null,"host_organization_lineage":[],"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"},"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":{"Attention":[0],"mechanisms":[1],"are":[2,117,177,184],"central":[3],"to":[4,14,23,61,139,180,206],"the":[5,30,49,52,69,76,80,84,89,100,105,120,124,153,157,161,170,181,211],"success":[6],"of":[7,88,94,104],"large":[8],"language":[9],"models":[10],"(LLMs),":[11],"enabling":[12],"them":[13],"capture":[15],"intricate":[16],"token":[17,54,91,107,122,131,150,158,183],"dependencies":[18],"and":[19,55,92,98,201],"implicitly":[20],"assign":[21],"importance":[22,151],"each":[24],"token.":[25,163],"Recent":[26],"studies":[27],"have":[28],"revealed":[29],"sink":[31,53,90,106,121,162,171,182],"token,":[32,172],"which":[33,156],"receives":[34],"disproportionately":[35],"high":[36],"attention":[37,60],"despite":[38],"their":[39,63],"limited":[40],"semantic":[41],"role.":[42],"In":[43],"this":[44],"paper,":[45],"we":[46,127,148,192],"first":[47],"expand":[48],"relationship":[50],"between":[51,83],"other":[56,95,114],"tokens,":[57],"moving":[58],"beyond":[59],"explore":[62],"similarity":[64,82],"in":[65,144,198],"hidden":[66,86,102],"states,":[67],"considering":[68],"layer":[70,207],"depth.":[71],"We":[72],"observe":[73],"that":[74,99,113,174,176,194],"as":[75],"layers":[77],"get":[78],"deeper,":[79],"cosine":[81],"normalized":[85,101],"states":[87,103],"those":[93],"tokens":[96,115,175],"increases,":[97],"exhibit":[108],"negligible":[109],"changes.":[110],"These":[111],"imply":[112],"consistently":[116],"directed":[118],"toward":[119,160],"throughout":[123],"layers.":[125],"Next,":[126],"propose":[128],"a":[129,145],"dynamic":[130],"selection":[132],"method,":[133],"called":[134],"OrthoRank,":[135],"using":[136],"these":[137],"findings":[138],"select":[140],"important":[141],"tokens.":[142],"Specifically,":[143],"certain":[146],"layer,":[147],"define":[149],"by":[152],"speed":[154],"at":[155,210],"moves":[159],"This":[164],"is":[165],"converted":[166],"into":[167],"orthogonality":[168],"with":[169,215],"meaning":[173],"more":[178],"orthogonal":[179],"assigned":[185],"greater":[186],"importance.":[187],"Finally,":[188],"through":[189],"extensive":[190],"experiments,":[191],"demonstrated":[193],"our":[195],"method":[196],"results":[197],"lower":[199],"perplexity":[200],"higher":[202],"zero-shot":[203],"accuracy":[204],"compared":[205],"pruning":[208],"methods":[209],"same":[212],"sparsity":[213],"ratio":[214],"comparable":[216],"throughput,":[217],"while":[218],"also":[219],"achieving":[220],"superior":[221],"performance":[222],"on":[223],"LongBench.":[224]},"counts_by_year":[],"updated_date":"2026-04-21T08:09:41.155169","created_date":"2025-10-20T00:00:00"}