{"id":"https://openalex.org/W7161683102","doi":"https://doi.org/10.48550/arxiv.2605.16928","title":"Full Attention Strikes Back: Transferring Full Attention into Sparse within Hundred Training Steps","display_name":"Full Attention Strikes Back: Transferring Full Attention into Sparse within Hundred Training Steps","publication_year":2026,"publication_date":"2026-05-16","ids":{"openalex":"https://openalex.org/W7161683102","doi":"https://doi.org/10.48550/arxiv.2605.16928"},"language":null,"primary_location":{"id":"doi:10.48550/arxiv.2605.16928","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2605.16928","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.2605.16928","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5136475870","display_name":"Yanke Zhou","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Zhou, Yanke","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5136495184","display_name":"Yiduo Li","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Li, Yiduo","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5089829380","display_name":"Hanlin Tang","orcid":"https://orcid.org/0000-0003-0306-1637"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Tang, Hanlin","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5136490548","display_name":"Maohua Li","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Li, Maohua","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5136482129","display_name":"Kan Liu","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Liu, Kan","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5136468461","display_name":"Lan Tao","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Lan, Tao","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5136473324","display_name":"Lin Qu","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Qu, Lin","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5136502270","display_name":"Yuan Yao","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Yao, Yuan","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5136481977","display_name":"Xiaoxing Ma","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Ma, Xiaoxing","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/T10028","display_name":"Topic Modeling","score":0.4004000127315521,"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":0.4004000127315521,"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":0.13989999890327454,"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/T11714","display_name":"Multimodal Machine Learning Applications","score":0.12120000272989273,"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/inference","display_name":"Inference","score":0.7422000169754028},{"id":"https://openalex.org/keywords/speedup","display_name":"Speedup","score":0.7049000263214111},{"id":"https://openalex.org/keywords/security-token","display_name":"Security token","score":0.6344000101089478},{"id":"https://openalex.org/keywords/context","display_name":"Context (archaeology)","score":0.5252000093460083},{"id":"https://openalex.org/keywords/heuristic","display_name":"Heuristic","score":0.5059000253677368},{"id":"https://openalex.org/keywords/training","display_name":"Training (meteorology)","score":0.4690000116825104},{"id":"https://openalex.org/keywords/selection","display_name":"Selection (genetic algorithm)","score":0.4410000145435333},{"id":"https://openalex.org/keywords/language-model","display_name":"Language model","score":0.41589999198913574}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.7847999930381775},{"id":"https://openalex.org/C2776214188","wikidata":"https://www.wikidata.org/wiki/Q408386","display_name":"Inference","level":2,"score":0.7422000169754028},{"id":"https://openalex.org/C68339613","wikidata":"https://www.wikidata.org/wiki/Q1549489","display_name":"Speedup","level":2,"score":0.7049000263214111},{"id":"https://openalex.org/C48145219","wikidata":"https://www.wikidata.org/wiki/Q1335365","display_name":"Security token","level":2,"score":0.6344000101089478},{"id":"https://openalex.org/C2779343474","wikidata":"https://www.wikidata.org/wiki/Q3109175","display_name":"Context (archaeology)","level":2,"score":0.5252000093460083},{"id":"https://openalex.org/C173801870","wikidata":"https://www.wikidata.org/wiki/Q201413","display_name":"Heuristic","level":2,"score":0.5059000253677368},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.47530001401901245},{"id":"https://openalex.org/C2777211547","wikidata":"https://www.wikidata.org/wiki/Q17141490","display_name":"Training (meteorology)","level":2,"score":0.4690000116825104},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.4471000134944916},{"id":"https://openalex.org/C81917197","wikidata":"https://www.wikidata.org/wiki/Q628760","display_name":"Selection (genetic algorithm)","level":2,"score":0.4410000145435333},{"id":"https://openalex.org/C137293760","wikidata":"https://www.wikidata.org/wiki/Q3621696","display_name":"Language model","level":2,"score":0.41589999198913574},{"id":"https://openalex.org/C115537543","wikidata":"https://www.wikidata.org/wiki/Q165596","display_name":"Cache","level":2,"score":0.4065999984741211},{"id":"https://openalex.org/C129844170","wikidata":"https://www.wikidata.org/wiki/Q41299","display_name":"Quadratic equation","level":2,"score":0.36149999499320984},{"id":"https://openalex.org/C157764524","wikidata":"https://www.wikidata.org/wiki/Q1383412","display_name":"Throughput","level":3,"score":0.35850000381469727},{"id":"https://openalex.org/C56372850","wikidata":"https://www.wikidata.org/wiki/Q1050404","display_name":"Sparse matrix","level":3,"score":0.33660000562667847},{"id":"https://openalex.org/C51632099","wikidata":"https://www.wikidata.org/wiki/Q3985153","display_name":"Training set","level":2,"score":0.3278000056743622},{"id":"https://openalex.org/C127705205","wikidata":"https://www.wikidata.org/wiki/Q5748245","display_name":"Heuristics","level":2,"score":0.32690000534057617},{"id":"https://openalex.org/C2777472644","wikidata":"https://www.wikidata.org/wiki/Q16968992","display_name":"Approximate inference","level":3,"score":0.30799999833106995},{"id":"https://openalex.org/C46743427","wikidata":"https://www.wikidata.org/wiki/Q1341685","display_name":"Inference engine","level":3,"score":0.27149999141693115},{"id":"https://openalex.org/C125411270","wikidata":"https://www.wikidata.org/wiki/Q18653","display_name":"Encoding (memory)","level":2,"score":0.25609999895095825},{"id":"https://openalex.org/C80444323","wikidata":"https://www.wikidata.org/wiki/Q2878974","display_name":"Theoretical computer science","level":1,"score":0.25589999556541443}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.48550/arxiv.2605.16928","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2605.16928","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.2605.16928","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2605.16928","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":[],"awards":[],"funders":[],"has_content":{"grobid_xml":false,"pdf":false},"content_urls":null,"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Long-context":[0],"inference":[1,206],"in":[2],"large":[3],"language":[4],"models":[5,59],"is":[6,66,87,111],"bottlenecked":[7],"by":[8,90],"the":[9,107,133,152],"quadratic":[10],"cost":[11],"of":[12,76],"full":[13,81,134],"attention.":[14,149],"Existing":[15],"efficient":[16],"alternatives":[17],"often":[18],"rely":[19],"either":[20],"on":[21,26,68,125,167],"native":[22,216],"sparse":[23,51,58,148,205,217],"training":[24,36,164,213],"or":[25],"heuristic":[27],"token":[28,109,145],"eviction,":[29],"creating":[30],"an":[31],"undesirable":[32],"trade-off":[33],"among":[34],"efficiency,":[35],"cost,":[37],"and":[38,52,105,141,170,194],"accuracy.":[39],"In":[40],"this":[41],"work,":[42],"we":[43,128],"show":[44,173],"that":[45,174,203],"full-attention":[46,212],"LLMs":[47],"are":[48],"already":[49],"intrinsically":[50],"can":[53,207],"be":[54,98,208],"transformed":[55],"into":[56],"highly":[57],"with":[60,101,159],"only":[61,72,137,160],"minimal":[62],"adaptation.":[63],"Our":[64],"approach":[65],"built":[67],"three":[69],"observations:":[70],"(1)":[71],"a":[73,91,102,143,161,187,196],"small":[74],"subset":[75],"attention":[77],"heads":[78,140],"truly":[79],"requires":[80],"long-context":[82,168],"processing;":[83],"(2)":[84],"long-range":[85],"retrieval":[86,139],"governed":[88],"primarily":[89],"low-dimensional":[92],"subspace,":[93],"allowing":[94],"relevant":[95],"tokens":[96],"to":[97,186],"retrieved":[99],"efficiently":[100],"16-dimensional":[103],"indexer;":[104],"(3)":[106],"useful":[108],"budget":[110],"strongly":[112],"query-dependent,":[113],"making":[114],"dynamic":[115],"top-$p$":[116],"selection":[117],"more":[118],"suitable":[119],"than":[120],"fixed":[121],"top-$k$":[122],"sparsification.":[123],"Based":[124],"these":[126],"insights,":[127],"propose":[129],"RTPurbo,":[130],"which":[131],"retains":[132],"KV":[135],"cache":[136],"for":[138,147],"introduces":[142],"lightweight":[144],"indexer":[146],"By":[150],"exploiting":[151],"model's":[153],"intrinsic":[154],"sparsity,":[155],"RTPurbo":[156,175],"achieves":[157],"sparsification":[158],"few":[162],"hundred":[163],"steps.":[165],"Experiments":[166],"benchmarks":[169],"reasoning":[171],"tasks":[172],"preserves":[176],"near-lossless":[177],"accuracy":[178],"while":[179],"delivering":[180],"substantial":[181],"efficiency":[182],"gains,":[183],"including":[184],"up":[185],"9.36$\\times$":[188],"prefill":[189],"speedup":[190],"at":[191],"1M":[192],"context":[193],"about":[195],"2.01$\\times$":[197],"decode":[198],"speedup.":[199],"These":[200],"results":[201],"suggest":[202],"strong":[204],"obtained":[209],"from":[210],"standard":[211],"without":[214],"expensive":[215],"pretraining.":[218]},"counts_by_year":[],"updated_date":"2026-07-01T06:00:48.157686","created_date":"2026-05-20T00:00:00"}