{"id":"https://openalex.org/W7166831151","doi":"https://doi.org/10.18653/v1/2026.acl-industry.62","title":"Memory-Efficient Structured Backpropagation for On-Device LLM Fine-Tuning","display_name":"Memory-Efficient Structured Backpropagation for On-Device LLM Fine-Tuning","publication_year":2026,"publication_date":"2026-01-01","ids":{"openalex":"https://openalex.org/W7166831151","doi":"https://doi.org/10.18653/v1/2026.acl-industry.62"},"language":null,"primary_location":{"id":"doi:10.18653/v1/2026.acl-industry.62","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.acl-industry.62","pdf_url":"https://aclanthology.org/2026.acl-industry.62.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":"Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 6: Industry Track)","raw_type":"proceedings-article"},"type":"conference-paper","indexed_in":["crossref"],"open_access":{"is_oa":true,"oa_status":"gold","oa_url":"https://aclanthology.org/2026.acl-industry.62.pdf","any_repository_has_fulltext":null},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5018102890","display_name":"Juneyoung Park","orcid":null},"institutions":[{"id":"https://openalex.org/I4210132720","display_name":"Laboratoire d'Optique Appliqu\u00e9e","ror":"https://ror.org/03czns913","country_code":"FR","type":"facility","lineage":["https://openalex.org/I1294671590","https://openalex.org/I1294671590","https://openalex.org/I142476485","https://openalex.org/I201181511","https://openalex.org/I4210098836","https://openalex.org/I4210132720","https://openalex.org/I4210145102","https://openalex.org/I4210145102","https://openalex.org/I4405258625"]}],"countries":["FR"],"is_corresponding":false,"raw_author_name":"JuneYoung Park","raw_affiliation_strings":["OptAI Inc"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"OptAI Inc","institution_ids":["https://openalex.org/I4210132720"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5139838322","display_name":"Yuri Hong","orcid":null},"institutions":[{"id":"https://openalex.org/I4210132720","display_name":"Laboratoire d'Optique Appliqu\u00e9e","ror":"https://ror.org/03czns913","country_code":"FR","type":"facility","lineage":["https://openalex.org/I1294671590","https://openalex.org/I1294671590","https://openalex.org/I142476485","https://openalex.org/I201181511","https://openalex.org/I4210098836","https://openalex.org/I4210132720","https://openalex.org/I4210145102","https://openalex.org/I4210145102","https://openalex.org/I4405258625"]}],"countries":["FR"],"is_corresponding":false,"raw_author_name":"Yuri Hong","raw_affiliation_strings":["OptAI Inc"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"OptAI Inc","institution_ids":["https://openalex.org/I4210132720"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5139814388","display_name":"Seongwan Kim","orcid":null},"institutions":[{"id":"https://openalex.org/I4210132720","display_name":"Laboratoire d'Optique Appliqu\u00e9e","ror":"https://ror.org/03czns913","country_code":"FR","type":"facility","lineage":["https://openalex.org/I1294671590","https://openalex.org/I1294671590","https://openalex.org/I142476485","https://openalex.org/I201181511","https://openalex.org/I4210098836","https://openalex.org/I4210132720","https://openalex.org/I4210145102","https://openalex.org/I4210145102","https://openalex.org/I4405258625"]}],"countries":["FR"],"is_corresponding":false,"raw_author_name":"Seongwan Kim","raw_affiliation_strings":["OptAI Inc"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"OptAI Inc","institution_ids":["https://openalex.org/I4210132720"]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5139849384","display_name":"Jaeho Lee","orcid":null},"institutions":[{"id":"https://openalex.org/I4210132720","display_name":"Laboratoire d'Optique Appliqu\u00e9e","ror":"https://ror.org/03czns913","country_code":"FR","type":"facility","lineage":["https://openalex.org/I1294671590","https://openalex.org/I1294671590","https://openalex.org/I142476485","https://openalex.org/I201181511","https://openalex.org/I4210098836","https://openalex.org/I4210132720","https://openalex.org/I4210145102","https://openalex.org/I4210145102","https://openalex.org/I4405258625"]}],"countries":["FR"],"is_corresponding":false,"raw_author_name":"Jaeho Lee","raw_affiliation_strings":["OptAI Inc"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"OptAI Inc","institution_ids":["https://openalex.org/I4210132720"]}]}],"institutions":[],"countries_distinct_count":1,"institutions_distinct_count":1,"corresponding_author_ids":[],"corresponding_institution_ids":["https://openalex.org/I4210132720"],"apc_list":null,"apc_paid":null,"fwci":0.0,"has_fulltext":true,"cited_by_count":0,"citation_normalized_percentile":{"value":0.86348427,"is_in_top_1_percent":false,"is_in_top_10_percent":false},"cited_by_percentile_year":null,"biblio":{"volume":null,"issue":null,"first_page":"906","last_page":"916"},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T11417","display_name":"Advancements in PLL and VCO Technologies","score":0.2605000138282776,"subfield":{"id":"https://openalex.org/subfields/2208","display_name":"Electrical and Electronic 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/T11417","display_name":"Advancements in PLL and VCO Technologies","score":0.2605000138282776,"subfield":{"id":"https://openalex.org/subfields/2208","display_name":"Electrical and Electronic Engineering"},"field":{"id":"https://openalex.org/fields/22","display_name":"Engineering"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}},{"id":"https://openalex.org/T11522","display_name":"VLSI and FPGA Design Techniques","score":0.07859999686479568,"subfield":{"id":"https://openalex.org/subfields/2208","display_name":"Electrical and Electronic Engineering"},"field":{"id":"https://openalex.org/fields/22","display_name":"Engineering"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}},{"id":"https://openalex.org/T10232","display_name":"Optical Network Technologies","score":0.05999999865889549,"subfield":{"id":"https://openalex.org/subfields/2208","display_name":"Electrical and Electronic 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/backpropagation","display_name":"Backpropagation","score":0.616100013256073},{"id":"https://openalex.org/keywords/artificial-neural-network","display_name":"Artificial neural network","score":0.4406999945640564},{"id":"https://openalex.org/keywords/pattern-recognition","display_name":"Pattern recognition (psychology)","score":0.2833999991416931},{"id":"https://openalex.org/keywords/noise","display_name":"Noise (video)","score":0.24979999661445618},{"id":"https://openalex.org/keywords/feature","display_name":"Feature (linguistics)","score":0.2493000030517578}],"concepts":[{"id":"https://openalex.org/C155032097","wikidata":"https://www.wikidata.org/wiki/Q798503","display_name":"Backpropagation","level":3,"score":0.616100013256073},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.49390000104904175},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.4659000039100647},{"id":"https://openalex.org/C50644808","wikidata":"https://www.wikidata.org/wiki/Q192776","display_name":"Artificial neural network","level":2,"score":0.4406999945640564},{"id":"https://openalex.org/C153180895","wikidata":"https://www.wikidata.org/wiki/Q7148389","display_name":"Pattern recognition (psychology)","level":2,"score":0.2833999991416931},{"id":"https://openalex.org/C99498987","wikidata":"https://www.wikidata.org/wiki/Q2210247","display_name":"Noise (video)","level":3,"score":0.24979999661445618},{"id":"https://openalex.org/C2776401178","wikidata":"https://www.wikidata.org/wiki/Q12050496","display_name":"Feature (linguistics)","level":2,"score":0.2493000030517578},{"id":"https://openalex.org/C71924100","wikidata":"https://www.wikidata.org/wiki/Q11190","display_name":"Medicine","level":0,"score":0.23180000483989716},{"id":"https://openalex.org/C168167062","wikidata":"https://www.wikidata.org/wiki/Q1117970","display_name":"Component (thermodynamics)","level":2,"score":0.225600004196167},{"id":"https://openalex.org/C108583219","wikidata":"https://www.wikidata.org/wiki/Q197536","display_name":"Deep learning","level":2,"score":0.22499999403953552}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.18653/v1/2026.acl-industry.62","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.acl-industry.62","pdf_url":"https://aclanthology.org/2026.acl-industry.62.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":"Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 6: Industry Track)","raw_type":"proceedings-article"}],"best_oa_location":{"id":"doi:10.18653/v1/2026.acl-industry.62","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.acl-industry.62","pdf_url":"https://aclanthology.org/2026.acl-industry.62.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":"Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 6: Industry Track)","raw_type":"proceedings-article"},"sustainable_development_goals":[{"score":0.5372319221496582,"id":"https://metadata.un.org/sdg/7","display_name":"Affordable and clean energy"}],"awards":[],"funders":[],"has_content":{"pdf":true,"grobid_xml":true},"content_urls":{"pdf":"https://content.openalex.org/works/W7166831151.pdf","grobid_xml":"https://content.openalex.org/works/W7166831151.grobid-xml"},"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"On-device":[0],"fine-tuning":[1,136],"enables":[2],"privacypreserving":[3],"personalization":[4],"of":[5,147],"large":[6],"language":[7],"models,":[8],"but":[9],"mobile":[10],"devices":[11],"impose":[12],"severe":[13],"memory":[14,31,35,93,129],"constraints,":[15],"typically":[16],"6-12GB":[17],"shared":[18],"across":[19],"all":[20],"workloads.Existing":[21],"approaches":[22],"force":[23],"a":[24],"trade-off":[25],"between":[26],"exact":[27],"gradients":[28,119],"with":[29,36,117],"high":[30],"(MeBP)":[32],"and":[33],"low":[34],"noisy":[37],"estimates":[38,113],"(MeZO).We":[39],"propose":[40],"Memory-efficient":[41],"Structured":[42],"Backpropagation":[43],"(MeSP),":[44],"which":[45],"bridges":[46],"this":[47],"gap":[48],"by":[49],"manually":[50],"deriving":[51],"backward":[52,73],"passes":[53],"that":[54,62,110],"exploit":[55],"LoRA's":[56],"low-rank":[57],"structure.Our":[58],"key":[59],"insight":[60],"is":[61],"the":[63,85],"intermediate":[64],"projection":[65],"h":[66],"=":[67],"xA":[68],"can":[69],"be":[70],"recomputed":[71],"during":[72],"at":[74,151],"minimal":[75],"cost":[76],"since":[77],"rank":[78],"r":[79],"\u226a":[80],"d":[81],"in":[82],",":[83],"eliminating":[84],"need":[86],"to":[87,96,132],"store":[88],"it.MeSP":[89],"achieves":[90],"49%":[91],"average":[92],"reduction":[94],"compared":[95],"MeBP":[97],"on":[98,140],"Qwen2.5":[99],"models":[100],"(0.5B-3B)":[101],"while":[102],"computing":[103],"mathematically":[104],"identical":[105],"gradients.Our":[106],"analysis":[107],"also":[108],"reveals":[109],"MeZO's":[111],"gradient":[112],"show":[114],"near-zero":[115],"correlation":[116],"true":[118],"(cosine":[120],"similarity":[121],"\u22480.001),":[122],"explaining":[123],"its":[124],"slow":[125],"convergence.MeSP":[126],"reduces":[127],"peak":[128],"from":[130],"361MB":[131],"136MB":[133],"for":[134],"Qwen2.5-0.5B,enabling":[135],"scenarios":[137],"previously":[138],"infeasible":[139],"memory-constrained":[141],"devices.We":[142],"provide":[143],"an":[144],"example":[145],"implementation":[146],"our":[148],"proposed":[149],"method":[150],"https:":[152],"//github.com/crinex/acl-mesp":[153]},"counts_by_year":[],"updated_date":"2026-07-29T14:22:42.915294","created_date":"2026-07-02T00:00:00"}