{"id":"https://openalex.org/W4407208976","doi":"https://doi.org/10.48550/arxiv.2502.02770","title":"Twilight: Adaptive Attention Sparsity with Hierarchical Top-$p$ Pruning","display_name":"Twilight: Adaptive Attention Sparsity with Hierarchical Top-$p$ Pruning","publication_year":2025,"publication_date":"2025-02-04","ids":{"openalex":"https://openalex.org/W4407208976","doi":"https://doi.org/10.48550/arxiv.2502.02770"},"language":"en","primary_location":{"id":"pmh:oai:arXiv.org:2502.02770","is_oa":true,"landing_page_url":"http://arxiv.org/abs/2502.02770","pdf_url":"https://arxiv.org/pdf/2502.02770","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":"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"},"type":"preprint","indexed_in":["arxiv","datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://arxiv.org/pdf/2502.02770","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":null,"display_name":"Lin, Chaofan","orcid":null},"institutions":[],"countries":[],"is_corresponding":true,"raw_author_name":"Lin, Chaofan","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"middle","author":{"id":null,"display_name":"Tang, Jiaming","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Tang, Jiaming","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5022200545","display_name":"Shuo Yang","orcid":"https://orcid.org/0000-0001-7106-9169"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Yang, Shuo","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5000296995","display_name":"Hanshuo Wang","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Wang, Hanshuo","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5116176156","display_name":"Tian Tang","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Tang, Tian","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5116176157","display_name":"Boyu Tian","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Tian, Boyu","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5116176158","display_name":"Ion Stoica","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Stoica, Ion","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5070926896","display_name":"Song Han","orcid":"https://orcid.org/0000-0002-4186-7618"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Han, Song","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5116176159","display_name":"Mingyu Gao","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Gao, Mingyu","raw_affiliation_strings":[],"affiliations":[]}],"institutions":[],"countries_distinct_count":0,"institutions_distinct_count":9,"corresponding_author_ids":[],"corresponding_institution_ids":[],"apc_list":null,"apc_paid":null,"fwci":null,"has_fulltext":true,"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/T11902","display_name":"Intelligent Tutoring Systems and Adaptive Learning","score":0.9829000234603882,"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/T11902","display_name":"Intelligent Tutoring Systems and Adaptive Learning","score":0.9829000234603882,"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/T11122","display_name":"Online Learning and Analytics","score":0.9696999788284302,"subfield":{"id":"https://openalex.org/subfields/1706","display_name":"Computer Science Applications"},"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/T12101","display_name":"Advanced Bandit Algorithms Research","score":0.9315999746322632,"subfield":{"id":"https://openalex.org/subfields/1803","display_name":"Management Science and Operations Research"},"field":{"id":"https://openalex.org/fields/18","display_name":"Decision Sciences"},"domain":{"id":"https://openalex.org/domains/2","display_name":"Social Sciences"}}],"keywords":[{"id":"https://openalex.org/keywords/twilight","display_name":"Twilight","score":0.8494561910629272},{"id":"https://openalex.org/keywords/pruning","display_name":"Pruning","score":0.7247982025146484},{"id":"https://openalex.org/keywords/computer-science","display_name":"Computer science","score":0.3948659598827362},{"id":"https://openalex.org/keywords/biology","display_name":"Biology","score":0.1966281533241272},{"id":"https://openalex.org/keywords/horticulture","display_name":"Horticulture","score":0.15807810425758362},{"id":"https://openalex.org/keywords/physics","display_name":"Physics","score":0.10071814060211182}],"concepts":[{"id":"https://openalex.org/C2777298776","wikidata":"https://www.wikidata.org/wiki/Q164160","display_name":"Twilight","level":2,"score":0.8494561910629272},{"id":"https://openalex.org/C108010975","wikidata":"https://www.wikidata.org/wiki/Q500094","display_name":"Pruning","level":2,"score":0.7247982025146484},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.3948659598827362},{"id":"https://openalex.org/C86803240","wikidata":"https://www.wikidata.org/wiki/Q420","display_name":"Biology","level":0,"score":0.1966281533241272},{"id":"https://openalex.org/C144027150","wikidata":"https://www.wikidata.org/wiki/Q48803","display_name":"Horticulture","level":1,"score":0.15807810425758362},{"id":"https://openalex.org/C121332964","wikidata":"https://www.wikidata.org/wiki/Q413","display_name":"Physics","level":0,"score":0.10071814060211182},{"id":"https://openalex.org/C1276947","wikidata":"https://www.wikidata.org/wiki/Q333","display_name":"Astronomy","level":1,"score":0.0}],"mesh":[],"locations_count":2,"locations":[{"id":"pmh:oai:arXiv.org:2502.02770","is_oa":true,"landing_page_url":"http://arxiv.org/abs/2502.02770","pdf_url":"https://arxiv.org/pdf/2502.02770","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":"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.2502.02770","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2502.02770","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":"article"}],"best_oa_location":{"id":"pmh:oai:arXiv.org:2502.02770","is_oa":true,"landing_page_url":"http://arxiv.org/abs/2502.02770","pdf_url":"https://arxiv.org/pdf/2502.02770","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":"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":[],"awards":[],"funders":[{"id":"https://openalex.org/F4320307758","display_name":"Eli Lilly and Company","ror":"https://ror.org/01qat3289"},{"id":"https://openalex.org/F4320321697","display_name":"LOEWE Zentrum AdRIA","ror":"https://ror.org/05n911h24"}],"has_content":{"pdf":true,"grobid_xml":true},"content_urls":{"pdf":"https://content.openalex.org/works/W4407208976.pdf","grobid_xml":"https://content.openalex.org/works/W4407208976.grobid-xml"},"referenced_works_count":0,"referenced_works":[],"related_works":["https://openalex.org/W4391375266","https://openalex.org/W2899084033","https://openalex.org/W2748952813","https://openalex.org/W2341012202","https://openalex.org/W2329039841","https://openalex.org/W2606348389","https://openalex.org/W1621998790","https://openalex.org/W1653644159","https://openalex.org/W4320717050","https://openalex.org/W4241575677"],"abstract_inverted_index":{"Leveraging":[0],"attention":[1,22,77,99],"sparsity":[2,94],"to":[3,29,44,75,91,95,120],"accelerate":[4],"long-context":[5],"large":[6],"language":[7],"models":[8],"(LLMs)":[9],"has":[10],"been":[11],"a":[12,31,36,89],"hot":[13],"research":[14],"topic.":[15],"However,":[16],"current":[17],"algorithms":[18],"such":[19],"as":[20],"sparse":[21,76,98],"or":[23],"key-value":[24],"(KV)":[25],"cache":[26],"compression":[27],"tend":[28],"use":[30],"fixed":[32],"budget,":[33],"which":[34],"presents":[35],"significant":[37],"challenge":[38],"during":[39],"deployment":[40],"because":[41],"it":[42],"fails":[43],"account":[45],"for":[46],"the":[47,54],"dynamic":[48],"nature":[49],"of":[50,116],"real-world":[51],"scenarios,":[52],"where":[53],"optimal":[55],"balance":[56],"between":[57],"accuracy":[58],"and":[59,126],"efficiency":[60],"can":[61,78,110],"vary":[62],"greatly.":[63],"In":[64],"this":[65],"paper,":[66],"we":[67,86],"find":[68],"that":[69,108],"borrowing":[70],"top-$p$":[71],"sampling":[72],"(nucleus":[73],"sampling)":[74],"surprisingly":[79],"achieve":[80],"adaptive":[81,93],"budgeting.":[82],"Based":[83],"on":[84],"this,":[85],"propose":[87],"Twilight,":[88],"framework":[90],"bring":[92],"any":[96],"existing":[97],"algorithm":[100],"without":[101],"sacrificing":[102],"their":[103],"accuracy.":[104],"Empirical":[105],"results":[106],"show":[107],"Twilight":[109],"adaptively":[111],"prune":[112],"at":[113],"most":[114],"98%":[115],"redundant":[117],"tokens,":[118],"leading":[119],"$15.4\\times$":[121],"acceleration":[122,128],"in":[123,129,134],"self-attention":[124],"operations":[125],"$3.9\\times$":[127],"end-to-end":[130],"per":[131],"token":[132],"latency":[133],"long":[135],"context":[136],"LLM":[137],"decoding.":[138]},"counts_by_year":[],"updated_date":"2026-04-21T08:09:41.155169","created_date":"2025-10-10T00:00:00"}
