{"id":"https://openalex.org/W4414359566","doi":"https://doi.org/10.24963/ijcai.2025/760","title":"Theoretical Insights into Fine-Tuning Attention Mechanism: Generalization and Optimization","display_name":"Theoretical Insights into Fine-Tuning Attention Mechanism: Generalization and Optimization","publication_year":2025,"publication_date":"2025-09-01","ids":{"openalex":"https://openalex.org/W4414359566","doi":"https://doi.org/10.24963/ijcai.2025/760"},"language":"en","primary_location":{"id":"doi:10.24963/ijcai.2025/760","is_oa":false,"landing_page_url":"https://doi.org/10.24963/ijcai.2025/760","pdf_url":null,"source":null,"license":null,"license_id":null,"version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Proceedings of the Thirty-Fourth International Joint Conference on Artificial Intelligence","raw_type":"proceedings-article"},"type":"article","indexed_in":["crossref"],"open_access":{"is_oa":false,"oa_status":"closed","oa_url":null,"any_repository_has_fulltext":false},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5042872078","display_name":"Xin Yao","orcid":"https://orcid.org/0000-0002-3662-9658"},"institutions":[{"id":"https://openalex.org/I4210121370","display_name":"Next Generation Technology (United States)","ror":"https://ror.org/02hswyz47","country_code":"US","type":"company","lineage":["https://openalex.org/I4210121370"]},{"id":"https://openalex.org/I4210126249","display_name":"Search (Poland)","ror":"https://ror.org/032t23b63","country_code":"PL","type":"company","lineage":["https://openalex.org/I4210126249"]},{"id":"https://openalex.org/I4210166345","display_name":"Intelligent Models (United States)","ror":"https://ror.org/05te6yt75","country_code":"US","type":"company","lineage":["https://openalex.org/I4210166345"]},{"id":"https://openalex.org/I78988378","display_name":"Renmin University of China","ror":"https://ror.org/041pakw92","country_code":"CN","type":"education","lineage":["https://openalex.org/I78988378"]}],"countries":["CN","PL","US"],"is_corresponding":true,"raw_author_name":"Xinhao Yao","raw_affiliation_strings":["Beijing Key Laboratory of Research on Large Models and Intelligent Governance","Engineering Research Center of Next-Generation Intelligent Search and Recommendation, MOE","Gaoling School of Artificial Intelligence, Renmin University of China, Beijing, China"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Beijing Key Laboratory of Research on Large Models and Intelligent Governance","institution_ids":["https://openalex.org/I4210166345"]},{"raw_affiliation_string":"Engineering Research Center of Next-Generation Intelligent Search and Recommendation, MOE","institution_ids":["https://openalex.org/I4210121370","https://openalex.org/I4210126249"]},{"raw_affiliation_string":"Gaoling School of Artificial Intelligence, Renmin University of China, Beijing, China","institution_ids":["https://openalex.org/I78988378"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5010274969","display_name":"Hongjin Qian","orcid":"https://orcid.org/0000-0003-4011-5673"},"institutions":[{"id":"https://openalex.org/I4210100255","display_name":"Beijing Academy of Artificial Intelligence","ror":"https://ror.org/016a74861","country_code":"CN","type":"other","lineage":["https://openalex.org/I4210100255"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Hongjin Qian","raw_affiliation_strings":["Beijing Academy of Artificial Intelligence"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Beijing Academy of Artificial Intelligence","institution_ids":["https://openalex.org/I4210100255"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5103895382","display_name":"Xiaolin Hu","orcid":"https://orcid.org/0009-0002-5493-5779"},"institutions":[{"id":"https://openalex.org/I78988378","display_name":"Renmin University of China","ror":"https://ror.org/041pakw92","country_code":"CN","type":"education","lineage":["https://openalex.org/I78988378"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Xiaolin Hu","raw_affiliation_strings":["Gaoling School of Artificial Intelligence, Renmin University of China, Beijing, China"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Gaoling School of Artificial Intelligence, Renmin University of China, Beijing, China","institution_ids":["https://openalex.org/I78988378"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5051833851","display_name":"Gengze Xu","orcid":null},"institutions":[{"id":"https://openalex.org/I78988378","display_name":"Renmin University of China","ror":"https://ror.org/041pakw92","country_code":"CN","type":"education","lineage":["https://openalex.org/I78988378"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Gengze Xu","raw_affiliation_strings":["Gaoling School of Artificial Intelligence, Renmin University of China, Beijing, China"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Gaoling School of Artificial Intelligence, Renmin University of China, Beijing, China","institution_ids":["https://openalex.org/I78988378"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5100431792","display_name":"Wei Liu","orcid":"https://orcid.org/0000-0002-3865-8145"},"institutions":[{"id":"https://openalex.org/I862669128","display_name":"Xiaomi (China)","ror":"https://ror.org/029f7bn57","country_code":"CN","type":"company","lineage":["https://openalex.org/I862669128"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Wei Liu","raw_affiliation_strings":["XiaoMi"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"XiaoMi","institution_ids":["https://openalex.org/I862669128"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5054843960","display_name":"Jian Luan","orcid":"https://orcid.org/0000-0002-2383-226X"},"institutions":[{"id":"https://openalex.org/I862669128","display_name":"Xiaomi (China)","ror":"https://ror.org/029f7bn57","country_code":"CN","type":"company","lineage":["https://openalex.org/I862669128"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Jian Luan","raw_affiliation_strings":["XiaoMi"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"XiaoMi","institution_ids":["https://openalex.org/I862669128"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5015829772","display_name":"Bin Wang","orcid":"https://orcid.org/0000-0001-9760-8343"},"institutions":[{"id":"https://openalex.org/I862669128","display_name":"Xiaomi (China)","ror":"https://ror.org/029f7bn57","country_code":"CN","type":"company","lineage":["https://openalex.org/I862669128"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Bin Wang","raw_affiliation_strings":["XiaoMi"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"XiaoMi","institution_ids":["https://openalex.org/I862669128"]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5102013358","display_name":"Yong Liu","orcid":"https://orcid.org/0000-0003-4098-2339"},"institutions":[{"id":"https://openalex.org/I4210121370","display_name":"Next Generation Technology (United States)","ror":"https://ror.org/02hswyz47","country_code":"US","type":"company","lineage":["https://openalex.org/I4210121370"]},{"id":"https://openalex.org/I4210126249","display_name":"Search (Poland)","ror":"https://ror.org/032t23b63","country_code":"PL","type":"company","lineage":["https://openalex.org/I4210126249"]},{"id":"https://openalex.org/I4210166345","display_name":"Intelligent Models (United States)","ror":"https://ror.org/05te6yt75","country_code":"US","type":"company","lineage":["https://openalex.org/I4210166345"]},{"id":"https://openalex.org/I78988378","display_name":"Renmin University of China","ror":"https://ror.org/041pakw92","country_code":"CN","type":"education","lineage":["https://openalex.org/I78988378"]}],"countries":["CN","PL","US"],"is_corresponding":false,"raw_author_name":"Yong Liu","raw_affiliation_strings":["Beijing Key Laboratory of Research on Large Models and Intelligent Governance","Engineering Research Center of Next-Generation Intelligent Search and Recommendation, MOE","Gaoling School of Artificial Intelligence, Renmin University of China, Beijing, China"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Beijing Key Laboratory of Research on Large Models and Intelligent Governance","institution_ids":["https://openalex.org/I4210166345"]},{"raw_affiliation_string":"Engineering Research Center of Next-Generation Intelligent Search and Recommendation, MOE","institution_ids":["https://openalex.org/I4210121370","https://openalex.org/I4210126249"]},{"raw_affiliation_string":"Gaoling School of Artificial Intelligence, Renmin University of China, Beijing, China","institution_ids":["https://openalex.org/I78988378"]}]}],"institutions":[],"countries_distinct_count":3,"institutions_distinct_count":8,"corresponding_author_ids":["https://openalex.org/A5042872078"],"corresponding_institution_ids":["https://openalex.org/I4210121370","https://openalex.org/I4210126249","https://openalex.org/I4210166345","https://openalex.org/I78988378"],"apc_list":null,"apc_paid":null,"fwci":2.2969,"has_fulltext":false,"cited_by_count":2,"citation_normalized_percentile":{"value":0.89291988,"is_in_top_1_percent":false,"is_in_top_10_percent":false},"cited_by_percentile_year":{"min":91,"max":98},"biblio":{"volume":null,"issue":null,"first_page":"6830","last_page":"6838"},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T10429","display_name":"EEG and Brain-Computer Interfaces","score":0.2800000011920929,"subfield":{"id":"https://openalex.org/subfields/2805","display_name":"Cognitive Neuroscience"},"field":{"id":"https://openalex.org/fields/28","display_name":"Neuroscience"},"domain":{"id":"https://openalex.org/domains/1","display_name":"Life Sciences"}},"topics":[{"id":"https://openalex.org/T10429","display_name":"EEG and Brain-Computer Interfaces","score":0.2800000011920929,"subfield":{"id":"https://openalex.org/subfields/2805","display_name":"Cognitive Neuroscience"},"field":{"id":"https://openalex.org/fields/28","display_name":"Neuroscience"},"domain":{"id":"https://openalex.org/domains/1","display_name":"Life Sciences"}},{"id":"https://openalex.org/T10581","display_name":"Neural dynamics and brain function","score":0.271699994802475,"subfield":{"id":"https://openalex.org/subfields/2805","display_name":"Cognitive Neuroscience"},"field":{"id":"https://openalex.org/fields/28","display_name":"Neuroscience"},"domain":{"id":"https://openalex.org/domains/1","display_name":"Life Sciences"}},{"id":"https://openalex.org/T10320","display_name":"Neural Networks and Applications","score":0.27070000767707825,"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"}}],"keywords":[{"id":"https://openalex.org/keywords/generalization","display_name":"Generalization","score":0.6930999755859375},{"id":"https://openalex.org/keywords/benchmark","display_name":"Benchmark (surveying)","score":0.6040999889373779},{"id":"https://openalex.org/keywords/transformer","display_name":"Transformer","score":0.4706000089645386},{"id":"https://openalex.org/keywords/convergence","display_name":"Convergence (economics)","score":0.4390000104904175},{"id":"https://openalex.org/keywords/range","display_name":"Range (aeronautics)","score":0.40709999203681946},{"id":"https://openalex.org/keywords/matrix","display_name":"Matrix (chemical analysis)","score":0.4002000093460083}],"concepts":[{"id":"https://openalex.org/C177148314","wikidata":"https://www.wikidata.org/wiki/Q170084","display_name":"Generalization","level":2,"score":0.6930999755859375},{"id":"https://openalex.org/C185798385","wikidata":"https://www.wikidata.org/wiki/Q1161707","display_name":"Benchmark (surveying)","level":2,"score":0.6040999889373779},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.5796999931335449},{"id":"https://openalex.org/C66322947","wikidata":"https://www.wikidata.org/wiki/Q11658","display_name":"Transformer","level":3,"score":0.4706000089645386},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.439300000667572},{"id":"https://openalex.org/C2777303404","wikidata":"https://www.wikidata.org/wiki/Q759757","display_name":"Convergence (economics)","level":2,"score":0.4390000104904175},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.43160000443458557},{"id":"https://openalex.org/C204323151","wikidata":"https://www.wikidata.org/wiki/Q905424","display_name":"Range (aeronautics)","level":2,"score":0.40709999203681946},{"id":"https://openalex.org/C106487976","wikidata":"https://www.wikidata.org/wiki/Q685816","display_name":"Matrix (chemical analysis)","level":2,"score":0.4002000093460083},{"id":"https://openalex.org/C2987595161","wikidata":"https://www.wikidata.org/wiki/Q141495","display_name":"Optimization algorithm","level":2,"score":0.3578000068664551},{"id":"https://openalex.org/C11413529","wikidata":"https://www.wikidata.org/wiki/Q8366","display_name":"Algorithm","level":1,"score":0.3434999883174896},{"id":"https://openalex.org/C2776291640","wikidata":"https://www.wikidata.org/wiki/Q2912517","display_name":"Value (mathematics)","level":2,"score":0.32580000162124634},{"id":"https://openalex.org/C57869625","wikidata":"https://www.wikidata.org/wiki/Q1783502","display_name":"Rate of convergence","level":3,"score":0.29679998755455017},{"id":"https://openalex.org/C126255220","wikidata":"https://www.wikidata.org/wiki/Q141495","display_name":"Mathematical optimization","level":1,"score":0.2935999929904938},{"id":"https://openalex.org/C80444323","wikidata":"https://www.wikidata.org/wiki/Q2878974","display_name":"Theoretical computer science","level":1,"score":0.2833000123500824},{"id":"https://openalex.org/C26517878","wikidata":"https://www.wikidata.org/wiki/Q228039","display_name":"Key (lock)","level":2,"score":0.27549999952316284},{"id":"https://openalex.org/C89611455","wikidata":"https://www.wikidata.org/wiki/Q6804646","display_name":"Mechanism (biology)","level":2,"score":0.2718000113964081}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.24963/ijcai.2025/760","is_oa":false,"landing_page_url":"https://doi.org/10.24963/ijcai.2025/760","pdf_url":null,"source":null,"license":null,"license_id":null,"version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Proceedings of the Thirty-Fourth International Joint Conference on Artificial Intelligence","raw_type":"proceedings-article"}],"best_oa_location":null,"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":{"Large":[0],"Language":[1],"Models":[2],"(LLMs),":[3],"built":[4],"on":[5,169,190],"Transformer":[6],"architectures,":[7],"exhibit":[8],"remarkable":[9,37],"generalization":[10],"across":[11],"a":[12,150,174],"wide":[13],"range":[14],"of":[15,47,57,71,77,141,183,196],"tasks.":[16],"However,":[17],"fine-tuning":[18,46,78,117,179],"these":[19,147,170],"models":[20],"for":[21,154,209],"specific":[22],"tasks":[23],"remains":[24],"resource-intensive":[25],"due":[26],"to":[27,40,135,146,159],"their":[28],"extensive":[29],"parameterization.":[30],"In":[31],"this":[32,197],"paper,":[33],"we":[34,172],"explore":[35],"two":[36],"phenomena":[38],"related":[39],"the":[41,45,55,58,75,86,95,100,139,155,194,206],"attention":[42],"mechanism":[43],"during":[44],"LLMs":[48,215],"(where":[49],"Wq,":[50],"Wk,":[51,122],"and":[52,61,102,123,161,165,186,211],"Wv":[53,87,103,156],"denote":[54],"weights":[56],"query,":[59],"key,":[60],"value":[62],"layers,":[63],"respectively).":[64],"The":[65,125],"first":[66],"phenomenon,":[67,127],"termed":[68],"\u201cUnequal":[69],"Importance":[70],"Attention":[72],"Matrices\u201d,":[73],"highlights":[74],"impact":[76],"different":[79],"weight":[80],"matrices.":[81,148],"It":[82],"shows":[83],"that":[84,177],"optimizing":[85,94],"matrix":[88,157],"yields":[89],"significantly":[90],"better":[91,115],"performance":[92],"than":[93,116],"Wk":[96,162],"matrix.":[97],"Fine-tuning":[98],"only":[99],"Wq":[101,160],"matrices":[104,120],"is":[105],"computationally":[106],"efficient":[107],"while":[108],"delivering":[109],"results":[110,189],"comparable":[111],"to,":[112],"or":[113],"even":[114],"all":[118],"three":[119],"(Wq,":[121],"Wv).":[124],"second":[126],"\u201cAttention":[128],"Matrices":[129],"with":[130],"Customized":[131],"Learning":[132],"Rate":[133],"Lead":[134],"Better":[136],"Convergence\u201d,":[137],"emphasizes":[138],"importance":[140],"assigning":[142],"distinct":[143],"learning":[144,152],"rates":[145],"Specifically,":[149],"higher":[151],"rate":[153],"compared":[158],"accelerates":[163],"convergence":[164],"improves":[166,178],"performance.":[167],"Building":[168],"insights,":[171],"propose":[173],"new":[175],"strategy":[176],"efficiency":[180],"in":[181,214],"terms":[182],"both":[184],"storage":[185],"time.":[187],"Experimental":[188],"benchmark":[191],"datasets":[192],"validate":[193],"effectiveness":[195],"approach,":[198],"supporting":[199],"our":[200],"theoretical":[201,207],"findings.":[202],"Our":[203],"analysis":[204],"lays":[205],"groundwork":[208],"configuring":[210],"improving":[212],"algorithms":[213],"fine-tuning.":[216]},"counts_by_year":[{"year":2026,"cited_by_count":1},{"year":2025,"cited_by_count":1}],"updated_date":"2026-05-21T06:26:12.895304","created_date":"2025-10-10T00:00:00"}