{"id":"https://openalex.org/W4415308159","doi":"https://doi.org/10.18653/v1/2026.findings-acl.1926","title":"The Sparse Frontier: Sparse Attention Trade-offs in Transformer LLMs","display_name":"The Sparse Frontier: Sparse Attention Trade-offs in Transformer LLMs","publication_year":2026,"publication_date":"2026-01-01","ids":{"openalex":"https://openalex.org/W4415308159","doi":"https://doi.org/10.18653/v1/2026.findings-acl.1926"},"language":"en","primary_location":{"id":"doi:10.18653/v1/2026.findings-acl.1926","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.findings-acl.1926","pdf_url":"https://aclanthology.org/2026.findings-acl.1926.pdf","source":null,"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Findings of the Association for Computational Linguistics: ACL 2026","raw_type":"proceedings-article"},"type":"conference-paper","indexed_in":["arxiv","crossref","datacite"],"open_access":{"is_oa":true,"oa_status":"gold","oa_url":"https://aclanthology.org/2026.findings-acl.1926.pdf","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5004341503","display_name":"Piotr Nawrot","orcid":"https://orcid.org/0000-0002-1195-864X"},"institutions":[{"id":"https://openalex.org/I98677209","display_name":"University of Edinburgh","ror":"https://ror.org/01nrxwf90","country_code":"GB","type":"education","lineage":["https://openalex.org/I98677209"]}],"countries":["GB"],"is_corresponding":false,"raw_author_name":"Piotr Nawrot","raw_affiliation_strings":["University of Edinburgh"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"University of Edinburgh","institution_ids":["https://openalex.org/I98677209"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5052582289","display_name":"Robert K.Y. Li","orcid":"https://orcid.org/0000-0002-9415-7855"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Jianing Li","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":null,"display_name":"Renjie Huang","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Renjie Huang","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5037310413","display_name":"Sebastian Ruder","orcid":null},"institutions":[{"id":"https://openalex.org/I98677209","display_name":"University of Edinburgh","ror":"https://ror.org/01nrxwf90","country_code":"GB","type":"education","lineage":["https://openalex.org/I98677209"]}],"countries":["GB"],"is_corresponding":false,"raw_author_name":"Sebastian Ruder","raw_affiliation_strings":["University of Edinburgh"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"University of Edinburgh","institution_ids":["https://openalex.org/I98677209"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5098901827","display_name":"Kelly Marchisio","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Kelly Marchisio","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5014613113","display_name":"Edoardo Maria Ponti","orcid":"https://orcid.org/0000-0002-6308-1050"},"institutions":[{"id":"https://openalex.org/I98677209","display_name":"University of Edinburgh","ror":"https://ror.org/01nrxwf90","country_code":"GB","type":"education","lineage":["https://openalex.org/I98677209"]}],"countries":["GB"],"is_corresponding":false,"raw_author_name":"Edoardo Ponti","raw_affiliation_strings":["University of Edinburgh"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"University of Edinburgh","institution_ids":["https://openalex.org/I98677209"]}]}],"institutions":[],"countries_distinct_count":1,"institutions_distinct_count":1,"corresponding_author_ids":[],"corresponding_institution_ids":[],"apc_list":null,"apc_paid":null,"fwci":19.5355,"has_fulltext":true,"cited_by_count":1,"citation_normalized_percentile":{"value":0.96554234,"is_in_top_1_percent":false,"is_in_top_10_percent":true},"cited_by_percentile_year":{"min":93,"max":97},"biblio":{"volume":null,"issue":null,"first_page":"38667","last_page":"38701"},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T10551","display_name":"Scheduling and Optimization Algorithms","score":0.7131999731063843,"subfield":{"id":"https://openalex.org/subfields/2209","display_name":"Industrial and Manufacturing Engineering"},"field":{"id":"https://openalex.org/fields/22","display_name":"Engineering"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}},"topics":[{"id":"https://openalex.org/T10551","display_name":"Scheduling and Optimization Algorithms","score":0.7131999731063843,"subfield":{"id":"https://openalex.org/subfields/2209","display_name":"Industrial and Manufacturing Engineering"},"field":{"id":"https://openalex.org/fields/22","display_name":"Engineering"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}}],"keywords":[{"id":"https://openalex.org/keywords/sparse-matrix","display_name":"Sparse matrix","score":0.5620999932289124},{"id":"https://openalex.org/keywords/sparse-approximation","display_name":"Sparse approximation","score":0.5151000022888184},{"id":"https://openalex.org/keywords/transformer","display_name":"Transformer","score":0.5077000260353088},{"id":"https://openalex.org/keywords/code","display_name":"Code (set theory)","score":0.3422999978065491},{"id":"https://openalex.org/keywords/neural-coding","display_name":"Neural coding","score":0.33799999952316284}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.619700014591217},{"id":"https://openalex.org/C56372850","wikidata":"https://www.wikidata.org/wiki/Q1050404","display_name":"Sparse matrix","level":3,"score":0.5620999932289124},{"id":"https://openalex.org/C124066611","wikidata":"https://www.wikidata.org/wiki/Q28684319","display_name":"Sparse approximation","level":2,"score":0.5151000022888184},{"id":"https://openalex.org/C66322947","wikidata":"https://www.wikidata.org/wiki/Q11658","display_name":"Transformer","level":3,"score":0.5077000260353088},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.5038999915122986},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.39489999413490295},{"id":"https://openalex.org/C2776760102","wikidata":"https://www.wikidata.org/wiki/Q5139990","display_name":"Code (set theory)","level":3,"score":0.3422999978065491},{"id":"https://openalex.org/C77637269","wikidata":"https://www.wikidata.org/wiki/Q7002051","display_name":"Neural coding","level":2,"score":0.33799999952316284},{"id":"https://openalex.org/C137635306","wikidata":"https://www.wikidata.org/wiki/Q182667","display_name":"Pareto principle","level":2,"score":0.32760000228881836},{"id":"https://openalex.org/C124101348","wikidata":"https://www.wikidata.org/wiki/Q172491","display_name":"Data mining","level":1,"score":0.30640000104904175},{"id":"https://openalex.org/C2993807640","wikidata":"https://www.wikidata.org/wiki/Q103709453","display_name":"Attention network","level":2,"score":0.2669000029563904}],"mesh":[],"locations_count":3,"locations":[{"id":"doi:10.18653/v1/2026.findings-acl.1926","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.findings-acl.1926","pdf_url":"https://aclanthology.org/2026.findings-acl.1926.pdf","source":null,"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Findings of the Association for Computational Linguistics: ACL 2026","raw_type":"proceedings-article"},{"id":"pmh:oai:arXiv.org:2504.17768","is_oa":true,"landing_page_url":"https://arxiv.org/abs/2504.17768","pdf_url":"https://arxiv.org/pdf/2504.17768","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,"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.2504.17768","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2504.17768","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,"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:10.18653/v1/2026.findings-acl.1926","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.findings-acl.1926","pdf_url":"https://aclanthology.org/2026.findings-acl.1926.pdf","source":null,"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Findings of the Association for Computational Linguistics: ACL 2026","raw_type":"proceedings-article"},"sustainable_development_goals":[],"awards":[{"id":"https://openalex.org/G373773752","display_name":null,"funder_award_id":"EP/S022481/1","funder_id":"https://openalex.org/F4320314731","funder_display_name":"UK Research and Innovation"}],"funders":[{"id":"https://openalex.org/F4320314731","display_name":"UK Research and Innovation","ror":"https://ror.org/001aqnf71"}],"has_content":{"pdf":true,"grobid_xml":true},"content_urls":{"pdf":"https://content.openalex.org/works/W4415308159.pdf","grobid_xml":"https://content.openalex.org/works/W4415308159.grobid-xml"},"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Sparse":[0],"attention":[1,62,76,92,186],"offers":[2],"a":[3,79,143],"promising":[4],"strategy":[5],"to":[6,20,34,51,58,124],"extend":[7],"long-context":[8],"capabilities":[9],"in":[10,174],"Transformer":[11],"LLMs,":[12],"yet":[13],"its":[14],"efficiency-accuracy":[15],"trade-offs":[16],"remain":[17],"unclear":[18],"due":[19],"the":[21,30,70,106,111,126,131],"lack":[22,132],"of":[23,36,74,128,133],"comprehensive":[24],"evaluation.We":[25],"address":[26],"this":[27],"gap":[28],"with":[29],"largestscale":[31],"empirical":[32],"analysis":[33,85],"date":[35],"trainingfree":[37],"sparse":[38,75,91,96,134,185],"attention,":[39],"evaluating":[40],"six":[41],"methods":[42,77,113,173],"across":[43],"multiple":[44],"model":[45],"families":[46],"and":[47,54,130,148,160,187],"sizes,":[48],"sequences":[49,166],"up":[50,57],"128K":[52],"tokens,":[53],"sparsity":[55,139,162],"levels":[56],"0.95":[59],"(i.e.,":[60],"1/20":[61],"budget)":[63],"on":[64],"nine":[65],"diverse":[66],"tasks.We":[67],"first":[68],"organise":[69],"rapidly":[71],"evolving":[72],"landscape":[73],"into":[78,140],"taxonomy":[80],"along":[81],"four":[82],"design":[83],"axes.Our":[84],"then":[86],"yields":[87],"actionable":[88],"insights:":[89],"1)":[90],"is":[93,194],"effective:":[94],"larger":[95],"models":[97],"outperform":[98],"smaller":[99],"dense":[100],"ones":[101],"at":[102,196],"equivalent":[103],"cost,":[104],"improving":[105],"Pareto":[107],"frontier;":[108],"2)":[109],"for":[110,183,190],"training-free":[112],"we":[114],"study,":[115],"fine-grained":[116,138],"per-query":[117],"importance":[118],"estimation":[119,129],"during":[120,151],"prefilling":[121],"remains":[122],"impractical-due":[123],"both":[125],"cost":[127],"kernels":[135],"that":[136,171],"translate":[137],"wallclock":[141],"gains-forcing":[142],"task-dependent":[144],"choice":[145],"between":[146],"global-to-token":[147],"block-to-block":[149],"selection.Instead,":[150],"decoding,":[152],"token-topage":[153],"selection":[154],"becomes":[155],"feasible,":[156],"enabling":[157],"better":[158],"generalisation":[159],"higher":[161,168],"tolerance;":[163],"3)":[164],"longer":[165],"tolerate":[167],"sparsity,":[169],"suggesting":[170],"fixed-budget":[172],"production":[175],"are":[176],"suboptimal.Together,":[177],"these":[178],"findings":[179],"provide":[180],"practical":[181],"guidance":[182],"deploying":[184],"methodological":[188],"recommendations":[189],"future":[191],"evaluations.Our":[192],"code":[193],"available":[195],"https://github.com/":[197],"PiotrNawrot/sparse-frontier.":[198]},"counts_by_year":[{"year":2026,"cited_by_count":1}],"updated_date":"2026-08-26T07:47:46.906454","created_date":"2025-10-18T00:00:00"}