{"id":"https://openalex.org/W4415080452","doi":"https://doi.org/10.48550/arxiv.2502.18821","title":"CAMEx: Curvature-aware Merging of Experts","display_name":"CAMEx: Curvature-aware Merging of Experts","publication_year":2025,"publication_date":"2025-02-26","ids":{"openalex":"https://openalex.org/W4415080452","doi":"https://doi.org/10.48550/arxiv.2502.18821"},"language":"en","primary_location":{"id":"pmh:oai:arXiv.org:2502.18821","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2502.18821","pdf_url":"https://arxiv.org/pdf/2502.18821","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":"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/2502.18821","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5028741605","display_name":"Dung V. Nguyen","orcid":"https://orcid.org/0000-0001-9390-9990","observed_orcids":["https://orcid.org/0000-0001-9390-9990"]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Nguyen, Dung V.","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":null,"display_name":"Nguyen, Minh H.","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Nguyen, Minh H.","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":null,"display_name":"Nguyen, Luc Q.","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Nguyen, Luc Q.","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5111236250","display_name":"Rachel S. Y. Teo","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Teo, Rachel S. Y.","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5000352150","display_name":"Tan Minh Nguyen","orcid":"https://orcid.org/0000-0002-9986-7629","observed_orcids":["https://orcid.org/0000-0002-9986-7629"]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Nguyen, Tan M.","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":null,"display_name":"Tran, Linh Duy","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Tran, Linh Duy","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":0.5005000233650208,"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":0.5005000233650208,"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/T11307","display_name":"Domain Adaptation and Few-Shot Learning","score":0.32100000977516174,"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/T11612","display_name":"Stochastic Gradient Optimization Techniques","score":0.07209999859333038,"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/natural-gradient","display_name":"natural gradient","score":0.6869999766349792},{"id":"https://openalex.org/keywords/fisher-information-matrix","display_name":"Fisher information matrix","score":0.6700000166893005},{"id":"https://openalex.org/keywords/large-language-models","display_name":"large language models","score":0.5720000267028809},{"id":"https://openalex.org/keywords/pretraining","display_name":"pre-training","score":0.7329999804496765},{"id":"https://openalex.org/keywords/fine-tuning","display_name":"fine-tuning","score":0.9900000095367432},{"id":"https://openalex.org/keywords/memory-efficient-optimization","display_name":"memory-efficient optimization","score":0.5080000162124634},{"id":"https://openalex.org/keywords/generalization","display_name":"generalization","score":0.6259999871253967}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.8162999749183655},{"id":"https://openalex.org/C177148314","wikidata":"https://www.wikidata.org/wiki/Q170084","display_name":"Generalization","level":2,"score":0.6711999773979187},{"id":"https://openalex.org/C2779960059","wikidata":"https://www.wikidata.org/wiki/Q7113681","display_name":"Overhead (engineering)","level":2,"score":0.5246000289916992},{"id":"https://openalex.org/C80444323","wikidata":"https://www.wikidata.org/wiki/Q2878974","display_name":"Theoretical computer science","level":1,"score":0.461899995803833},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.42969998717308044},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.4090999960899353},{"id":"https://openalex.org/C2776359362","wikidata":"https://www.wikidata.org/wiki/Q2145286","display_name":"Representation (politics)","level":3,"score":0.3783999979496002},{"id":"https://openalex.org/C195324797","wikidata":"https://www.wikidata.org/wiki/Q33742","display_name":"Natural language","level":2,"score":0.373199999332428},{"id":"https://openalex.org/C124101348","wikidata":"https://www.wikidata.org/wiki/Q172491","display_name":"Data mining","level":1,"score":0.36390000581741333},{"id":"https://openalex.org/C26517878","wikidata":"https://www.wikidata.org/wiki/Q228039","display_name":"Key (lock)","level":2,"score":0.36149999499320984},{"id":"https://openalex.org/C2780385302","wikidata":"https://www.wikidata.org/wiki/Q367158","display_name":"Protocol (science)","level":3,"score":0.3562000095844269},{"id":"https://openalex.org/C206345919","wikidata":"https://www.wikidata.org/wiki/Q20380951","display_name":"Resource (disambiguation)","level":2,"score":0.35589998960494995},{"id":"https://openalex.org/C151201525","wikidata":"https://www.wikidata.org/wiki/Q177239","display_name":"Limit (mathematics)","level":2,"score":0.35420000553131104},{"id":"https://openalex.org/C2776760102","wikidata":"https://www.wikidata.org/wiki/Q5139990","display_name":"Code (set theory)","level":3,"score":0.34389999508857727},{"id":"https://openalex.org/C125411270","wikidata":"https://www.wikidata.org/wiki/Q18653","display_name":"Encoding (memory)","level":2,"score":0.3294999897480011},{"id":"https://openalex.org/C45374587","wikidata":"https://www.wikidata.org/wiki/Q12525525","display_name":"Computation","level":2,"score":0.3294000029563904},{"id":"https://openalex.org/C159620131","wikidata":"https://www.wikidata.org/wiki/Q1938983","display_name":"Spatial analysis","level":2,"score":0.3174000084400177},{"id":"https://openalex.org/C129782007","wikidata":"https://www.wikidata.org/wiki/Q162886","display_name":"Euclidean geometry","level":2,"score":0.3052999973297119},{"id":"https://openalex.org/C29406490","wikidata":"https://www.wikidata.org/wiki/Q1420659","display_name":"Fisher information","level":2,"score":0.3034000098705292},{"id":"https://openalex.org/C127964446","wikidata":"https://www.wikidata.org/wiki/Q1092142","display_name":"Computational resource","level":3,"score":0.298799991607666},{"id":"https://openalex.org/C43126263","wikidata":"https://www.wikidata.org/wiki/Q128751","display_name":"Source code","level":2,"score":0.2939999997615814},{"id":"https://openalex.org/C11413529","wikidata":"https://www.wikidata.org/wiki/Q8366","display_name":"Algorithm","level":1,"score":0.2912999987602234},{"id":"https://openalex.org/C99844830","wikidata":"https://www.wikidata.org/wiki/Q102441924","display_name":"Scaling","level":2,"score":0.27889999747276306},{"id":"https://openalex.org/C2779439875","wikidata":"https://www.wikidata.org/wiki/Q1078276","display_name":"Natural language understanding","level":3,"score":0.27160000801086426},{"id":"https://openalex.org/C195065555","wikidata":"https://www.wikidata.org/wiki/Q214881","display_name":"Curvature","level":2,"score":0.25209999084472656}],"mesh":[],"locations_count":3,"locations":[{"id":"pmh:oai:arXiv.org:2502.18821","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2502.18821","pdf_url":"https://arxiv.org/pdf/2502.18821","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":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},{"id":"pmh:oai:scholarbank.nus.edu.sg:10635/321196","doi":null,"is_oa":false,"landing_page_url":"https://scholarbank.nus.edu.sg/handle/10635/321196","pdf_url":null,"source":{"id":"https://openalex.org/S7407052290","display_name":"National University of Singapore","issn_l":null,"issn":null,"is_oa":false,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":null,"host_organization_name":null,"host_organization_lineage":[],"host_organization_lineage_names":[],"type":"repository"},"license":null,"license_id":null,"version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":"Elements","raw_type":"Conference Paper"},{"id":"doi:10.48550/arxiv.2502.18821","doi":"https://doi.org/10.48550/arxiv.2502.18821","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2502.18821","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":"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":"pmh:oai:arXiv.org:2502.18821","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2502.18821","pdf_url":"https://arxiv.org/pdf/2502.18821","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":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},"sustainable_development_goals":[],"sustainable_development_goals_aurora":[],"study_designs":[],"awards":[],"funders":[],"has_content":{"pdf":false,"grobid_xml":false},"content_urls":null,"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Existing":[0],"methods":[1,45],"for":[2,83],"merging":[3,44,75,151,172],"experts":[4],"during":[5,30,163],"model":[6,109],"training":[7],"and":[8,50,111,120,127,165,192,198],"fine-tuning":[9],"predominantly":[10],"rely":[11],"on":[12],"Euclidean":[13],"geometry,":[14],"which":[15],"assumes":[16],"a":[17,72,170],"flat":[18],"parameter":[19,36,89,104],"space.":[20],"This":[21,115],"assumption":[22],"can":[23],"limit":[24],"the":[25,31,35,55,84,88,100,103,112,131,203],"model's":[26],"generalization":[27,129],"ability,":[28],"especially":[29],"pre-training":[32,119,164],"phase,":[33],"where":[34],"manifold":[37],"might":[38],"exhibit":[39],"more":[40,97],"complex":[41],"curvature.":[42],"Curvature-aware":[43],"typically":[46,135],"require":[47],"additional":[48],"information":[49],"computational":[51,183],"resources":[52],"to":[53,81,99,160,201],"approximate":[54],"Fisher":[56],"Information":[57],"Matrix,":[58],"adding":[59],"memory":[60,133],"overhead.":[61],"In":[62],"this":[63],"paper,":[64],"we":[65,168,194],"introduce":[66,169],"CAMEx":[67,95,145],"(Curvature-Aware":[68],"Merging":[69],"of":[70,87,102,188,205],"Experts),":[71],"novel":[73],"expert":[74,150],"protocol":[76],"that":[77,174],"incorporates":[78],"natural":[79,93,155],"gradients":[80],"account":[82],"non-Euclidean":[85],"curvature":[86],"manifold.":[90],"By":[91],"leveraging":[92],"gradients,":[94],"adapts":[96],"effectively":[98],"structure":[101],"space,":[105],"improving":[106],"alignment":[107],"between":[108],"updates":[110],"manifold's":[113],"geometry.":[114],"approach":[116],"enhances":[117],"both":[118,196],"fine-tuning,":[121],"resulting":[122],"in":[123],"better":[124],"optimization":[125],"trajectories":[126],"improved":[128],"without":[130],"substantial":[132],"overhead":[134],"associated":[136],"with":[137],"curvature-aware":[138],"methods.":[139],"Our":[140],"contributions":[141],"are":[142],"threefold:":[143],"(1)":[144],"significantly":[146],"outperforms":[147],"traditional":[148],"Euclidean-based":[149],"techniques":[152],"across":[153],"various":[154],"language":[156,190],"processing":[157],"tasks,":[158],"leading":[159],"enhanced":[161],"performance":[162,180],"fine-tuning;":[166],"(2)":[167],"dynamic":[171],"architecture":[173],"optimizes":[175],"resource":[176],"utilization,":[177],"achieving":[178],"high":[179],"while":[181],"reducing":[182],"costs,":[184],"facilitating":[185],"efficient":[186],"scaling":[187],"large":[189],"models;":[191],"(3)":[193],"provide":[195],"theoretical":[197],"empirical":[199],"evidence":[200],"demonstrate":[202],"efficiency":[204],"our":[206],"proposed":[207],"method.":[208],"The":[209],"code":[210],"is":[211],"publicly":[212],"available":[213],"at:":[214],"https://github.com/kpup1710/CAMEx.":[215]},"counts_by_year":[],"updated_date":"2026-09-29T14:09:16.559479","created_date":"2025-10-12T00:00:00"}