{"id":"https://openalex.org/W7214737102","doi":"https://doi.org/10.48550/arxiv.2609.34650","title":"When Can Attention Heads Be Statically Defined?","display_name":"When Can Attention Heads Be Statically Defined?","publication_year":2026,"publication_date":"2026-09-28","ids":{"openalex":"https://openalex.org/W7214737102","doi":"https://doi.org/10.48550/arxiv.2609.34650"},"language":"en","primary_location":{"id":"pmh:oai:arXiv.org:2609.34650","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2609.34650","pdf_url":"https://arxiv.org/pdf/2609.34650","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-sa","license_id":"https://openalex.org/licenses/cc-by-sa","version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},"volume":null,"type":"preprint","indexed_in":["arxiv","datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://arxiv.org/pdf/2609.34650","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5009974899","display_name":"Weixian Waylon Li","orcid":"https://orcid.org/0000-0002-4196-9462","observed_orcids":["https://orcid.org/0000-0002-4196-9462"]},"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":"Li, Weixian Waylon","raw_affiliation_strings":["University of Edinburgh , United Kingdom"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"University of Edinburgh , United Kingdom","institution_ids":["https://openalex.org/I98677209"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5056480393","display_name":"Yintao Tai","orcid":null,"observed_orcids":[]},"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":"Tai, Yintao","raw_affiliation_strings":["University of Edinburgh , United Kingdom"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"University of Edinburgh , United Kingdom","institution_ids":["https://openalex.org/I98677209"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5005636620","display_name":"Marcio Fonseca","orcid":"https://orcid.org/0009-0002-9340-0793","observed_orcids":["https://orcid.org/0009-0002-9340-0793"]},"institutions":[{"id":"https://openalex.org/I4405263582","display_name":"Chamber of Deputies - Brazil","ror":"https://ror.org/033av2692","country_code":"BR","type":"government","lineage":["https://openalex.org/I4405263582"]}],"countries":["BR"],"is_corresponding":false,"raw_author_name":"Fonseca, Marcio","raw_affiliation_strings":["Chamber of Deputies , Brazil"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Chamber of Deputies , Brazil","institution_ids":["https://openalex.org/I4405263582"]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5030503109","display_name":"Shay B. Cohen","orcid":"https://orcid.org/0000-0003-4753-8353","observed_orcids":["https://orcid.org/0000-0003-4753-8353"]},"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":"Cohen, Shay B.","raw_affiliation_strings":["University of Edinburgh , United Kingdom"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"University of Edinburgh , United Kingdom","institution_ids":["https://openalex.org/I98677209"]}]}],"institutions":[],"countries_distinct_count":2,"institutions_distinct_count":2,"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/T10028","display_name":"Topic Modeling","score":0.7939000129699707,"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/T10028","display_name":"Topic Modeling","score":0.7939000129699707,"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/T10181","display_name":"Natural Language Processing Techniques","score":0.07169999927282333,"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/T10036","display_name":"Advanced Neural Network Applications","score":0.06270000338554382,"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/training-cost-reduction","display_name":"training cost reduction","score":0.847000002861023},{"id":"https://openalex.org/keywords/associative-recall","display_name":"associative recall","score":0.8180000185966492},{"id":"https://openalex.org/keywords/attention-head-pruning","display_name":"attention head pruning","score":0.6060000061988831},{"id":"https://openalex.org/keywords/large-language-models","display_name":"large language models","score":0.527999997138977}],"concepts":[{"id":"https://openalex.org/C100279451","wikidata":"https://www.wikidata.org/wiki/Q372193","display_name":"Perplexity","level":3,"score":0.9487000107765198},{"id":"https://openalex.org/C108010975","wikidata":"https://www.wikidata.org/wiki/Q500094","display_name":"Pruning","level":2,"score":0.7581999897956848},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.7303000092506409},{"id":"https://openalex.org/C129844170","wikidata":"https://www.wikidata.org/wiki/Q41299","display_name":"Quadratic equation","level":2,"score":0.6171000003814697},{"id":"https://openalex.org/C45374587","wikidata":"https://www.wikidata.org/wiki/Q12525525","display_name":"Computation","level":2,"score":0.569100022315979},{"id":"https://openalex.org/C2778112365","wikidata":"https://www.wikidata.org/wiki/Q3511065","display_name":"Sequence (biology)","level":2,"score":0.557699978351593},{"id":"https://openalex.org/C196083921","wikidata":"https://www.wikidata.org/wiki/Q7915758","display_name":"Variance (accounting)","level":2,"score":0.5440000295639038},{"id":"https://openalex.org/C206588197","wikidata":"https://www.wikidata.org/wiki/Q846574","display_name":"Reuse","level":2,"score":0.5354999899864197},{"id":"https://openalex.org/C74193536","wikidata":"https://www.wikidata.org/wiki/Q574844","display_name":"Kernel (algebra)","level":2,"score":0.5299999713897705},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.46790000796318054},{"id":"https://openalex.org/C11413529","wikidata":"https://www.wikidata.org/wiki/Q8366","display_name":"Algorithm","level":1,"score":0.43860000371932983},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.3483999967575073},{"id":"https://openalex.org/C80444323","wikidata":"https://www.wikidata.org/wiki/Q2878974","display_name":"Theoretical computer science","level":1,"score":0.32670000195503235},{"id":"https://openalex.org/C3020493868","wikidata":"https://www.wikidata.org/wiki/Q55631277","display_name":"Real world data","level":2,"score":0.31139999628067017},{"id":"https://openalex.org/C51632099","wikidata":"https://www.wikidata.org/wiki/Q3985153","display_name":"Training set","level":2,"score":0.3107999861240387},{"id":"https://openalex.org/C137293760","wikidata":"https://www.wikidata.org/wiki/Q3621696","display_name":"Language model","level":2,"score":0.2897000014781952},{"id":"https://openalex.org/C157764524","wikidata":"https://www.wikidata.org/wiki/Q1383412","display_name":"Throughput","level":3,"score":0.26170000433921814},{"id":"https://openalex.org/C90119067","wikidata":"https://www.wikidata.org/wiki/Q43260","display_name":"Polynomial","level":2,"score":0.2590000033378601},{"id":"https://openalex.org/C2779662365","wikidata":"https://www.wikidata.org/wiki/Q5416694","display_name":"Event (particle physics)","level":2,"score":0.2581999897956848},{"id":"https://openalex.org/C311688","wikidata":"https://www.wikidata.org/wiki/Q2393193","display_name":"Time complexity","level":2,"score":0.2540999948978424}],"mesh":[],"locations_count":2,"locations":[{"id":"pmh:oai:arXiv.org:2609.34650","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2609.34650","pdf_url":"https://arxiv.org/pdf/2609.34650","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-sa","license_id":"https://openalex.org/licenses/cc-by-sa","version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},{"id":"doi:10.48550/arxiv.2609.34650","doi":"https://doi.org/10.48550/arxiv.2609.34650","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2609.34650","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":null,"license_id":null,"version":null,"is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"Preprint"}],"best_oa_location":{"id":"pmh:oai:arXiv.org:2609.34650","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2609.34650","pdf_url":"https://arxiv.org/pdf/2609.34650","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-sa","license_id":"https://openalex.org/licenses/cc-by-sa","version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},"sustainable_development_goals":[],"study_designs":[],"awards":[{"id":"https://openalex.org/G2912470688","display_name":null,"funder_award_id":"ST/AIRR/I-A-I/1023","funder_id":"https://openalex.org/F4320314731","funder_display_name":"UK Research and Innovation"},{"id":"https://openalex.org/G963556061","display_name":null,"funder_award_id":"ST/AIRR/I-A-I/1023","funder_id":"https://openalex.org/F4320334632","funder_display_name":"Science and Technology Facilities Council"}],"funders":[{"id":"https://openalex.org/F4320314707","display_name":"Government of the United Kingdom","ror":"https://ror.org/05wnh3t63"},{"id":"https://openalex.org/F4320314731","display_name":"UK Research and Innovation","ror":"https://ror.org/001aqnf71"},{"id":"https://openalex.org/F4320320360","display_name":"University of Bristol","ror":"https://ror.org/0524sp257"},{"id":"https://openalex.org/F4320334632","display_name":"Science and Technology Facilities Council","ror":"https://ror.org/057g20z61"}],"has_content":{"grobid_xml":true,"pdf":true},"content_urls":{"pdf":"https://content.openalex.org/works/W7214737102.pdf","grobid_xml":"https://content.openalex.org/works/W7214737102.grobid-xml"},"referenced_works_count":0,"referenced_works":[],"related_works":["https://openalex.org/W7151232163","https://openalex.org/W4414586697","https://openalex.org/W7202147293","https://openalex.org/W7160610055","https://openalex.org/W7161842911","https://openalex.org/W7155244598","https://openalex.org/W7165640989","https://openalex.org/W4406026730","https://openalex.org/W7128492566","https://openalex.org/W4415337169"],"abstract_inverted_index":{"Some":[0],"attention":[1,43,91,161],"heads":[2,35,82,92],"learn":[3],"similar":[4],"patterns":[5,10,56,76],"across":[6],"inputs.":[7],"Reusing":[8],"these":[9,54],"could":[11],"reduce":[12],"training":[13],"cost":[14],"by":[15],"avoiding":[16],"repeated":[17],"query-key":[18],"score":[19],"computation":[20],"and":[21,40,59,77,80,102,112,137,162],"softmax.":[22],"Through":[23],"controlled":[24],"pretraining":[25],"comparisons,":[26],"we":[27],"identify":[28],"Selective":[29],"Attention":[30],"Freezing":[31],"(SAF),":[32],"which":[33],"selects":[34],"with":[36,45,57,105,125,146],"low":[37],"attention-pattern":[38],"variance":[39],"replaces":[41],"their":[42],"weights":[44],"fitted":[46],"post-softmax":[47],"means":[48],"halfway":[49],"through":[50],"training.":[51],"We":[52],"represent":[53],"fixed":[55,156],"absolute-position":[58],"relative-distance":[60],"preferences,":[61],"reducing":[62],"storage":[63],"from":[64],"quadratic":[65],"to":[66,150],"linear":[67],"in":[68],"sequence":[69],"length.":[70],"A":[71],"fused":[72],"kernel":[73],"reconstructs":[74],"the":[75,143],"executes":[78],"ordinary-attention":[79],"replaced":[81],"together.":[83],"At":[84,110],"matched":[85],"training-token":[86],"budgets,":[87],"replacing":[88],"25%":[89,147],"of":[90],"gives":[93],"1.056x":[94],"faster":[95,119],"post-replacement":[96,115],"optimiser":[97],"updates":[98,116],"at":[99,154],"124M":[100,144],"parameters":[101],"4K":[103],"context,":[104,114],"a":[106,126,155],"0.77%":[107],"perplexity":[108,128],"increase.":[109,129],"1B":[111],"8K":[113],"are":[117],"1.068x":[118],"on":[120],"four":[121],"GPUs":[122],"including":[123],"communication,":[124],"0.51%":[127],"The":[130],"resulting":[131],"models":[132],"also":[133],"accelerate":[134],"long-input":[135],"finetuning":[136],"causal":[138],"prefill.":[139],"After":[140],"associative-recall":[141],"adaptation,":[142],"model":[145],"replacement":[148],"generalises":[149],"more":[151],"key-value":[152],"pairs":[153],"length":[157],"better":[158],"than":[159],"ordinary":[160],"two":[163],"pruning":[164],"controls.":[165]},"counts_by_year":[],"updated_date":"2026-10-28T10:33:24.094632","created_date":"2026-09-30T00:00:00"}