{"id":"https://openalex.org/W7214807286","doi":"https://doi.org/10.48550/arxiv.2609.35568","title":"From Experience to Expertise: Adoption-Aware Memory Learning for Data-Scarce NPU Kernel Synthesis","display_name":"From Experience to Expertise: Adoption-Aware Memory Learning for Data-Scarce NPU Kernel Synthesis","publication_year":2026,"publication_date":"2026-09-28","ids":{"openalex":"https://openalex.org/W7214807286","doi":"https://doi.org/10.48550/arxiv.2609.35568"},"language":"en","primary_location":{"id":"pmh:oai:arXiv.org:2609.35568","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2609.35568","pdf_url":"https://arxiv.org/pdf/2609.35568","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","is_traditional":true,"indexed_in":["arxiv","datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://arxiv.org/pdf/2609.35568","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5140357828","display_name":"Longxiao Fan","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I24943067","display_name":"Fudan University","ror":"https://ror.org/013q1eq08","country_code":"CN","type":"education","lineage":["https://openalex.org/I24943067"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Fan, Longxiao","raw_affiliation_strings":["Fudan University"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Fudan University","institution_ids":["https://openalex.org/I24943067"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5153504323","display_name":"Tao Zhang","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I126520041","display_name":"University of Science and Technology of China","ror":"https://ror.org/04c4dkn09","country_code":"CN","type":"education","lineage":["https://openalex.org/I126520041","https://openalex.org/I19820366"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Zhang, Tao","raw_affiliation_strings":["University of Science and Technology of China"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"University of Science and Technology of China","institution_ids":["https://openalex.org/I126520041"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5153482999","display_name":"Han Yan","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I24943067","display_name":"Fudan University","ror":"https://ror.org/013q1eq08","country_code":"CN","type":"education","lineage":["https://openalex.org/I24943067"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Yan, Han","raw_affiliation_strings":["Fudan University"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Fudan University","institution_ids":["https://openalex.org/I24943067"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5153405027","display_name":"Jiajun Li","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Li, Jiajun","raw_affiliation_strings":["Huawei Technologies Ltd"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Huawei Technologies Ltd","institution_ids":[]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5000111354","display_name":"Mingcong Song","orcid":"https://orcid.org/0009-0002-5289-685X","observed_orcids":["https://orcid.org/0009-0002-5289-685X"]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Song, Mingcong","raw_affiliation_strings":["Huawei Technologies Ltd"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Huawei Technologies Ltd","institution_ids":[]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5103437255","display_name":"Guoping Long","orcid":"https://orcid.org/0009-0006-3176-7572","observed_orcids":["https://orcid.org/0009-0006-3176-7572"]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Long, Guoping","raw_affiliation_strings":["Huawei Technologies Ltd"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Huawei Technologies Ltd","institution_ids":[]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5004222682","display_name":"Hongjie Si","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Si, Hongjie","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5153427769","display_name":"Weiwei Sun","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I24943067","display_name":"Fudan University","ror":"https://ror.org/013q1eq08","country_code":"CN","type":"education","lineage":["https://openalex.org/I24943067"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Sun, Weiwei","raw_affiliation_strings":["Fudan University"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Fudan University","institution_ids":["https://openalex.org/I24943067"]}]}],"institutions":[],"countries_distinct_count":1,"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/T10054","display_name":"Parallel Computing and Optimization Techniques","score":0.9883000254631042,"subfield":{"id":"https://openalex.org/subfields/1708","display_name":"Hardware and Architecture"},"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/T10054","display_name":"Parallel Computing and Optimization Techniques","score":0.9883000254631042,"subfield":{"id":"https://openalex.org/subfields/1708","display_name":"Hardware and Architecture"},"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/T10904","display_name":"Embedded Systems Design Techniques","score":0.006599999964237213,"subfield":{"id":"https://openalex.org/subfields/1708","display_name":"Hardware and Architecture"},"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/T10260","display_name":"Software Engineering Research","score":0.0010000000474974513,"subfield":{"id":"https://openalex.org/subfields/1710","display_name":"Information Systems"},"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/llm-coding-agents","display_name":"LLM coding agents","score":0.6729999780654907},{"id":"https://openalex.org/keywords/domain-specific-architectures","display_name":"domain-specific architectures","score":0.6629999876022339},{"id":"https://openalex.org/keywords/secondary-storage","display_name":"external memory","score":0.550000011920929},{"id":"https://openalex.org/keywords/credit-assignment","display_name":"credit assignment","score":0.7409999966621399},{"id":"https://openalex.org/keywords/learning-and-memory","display_name":"learning and memory","score":0.4000000059604645},{"id":"https://openalex.org/keywords/knowledge-consolidation","display_name":"knowledge consolidation","score":0.4699999988079071},{"id":"https://openalex.org/keywords/memory-hierarchy","display_name":"memory hierarchy","score":0.4699999988079071}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.7641000151634216},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.5224000215530396},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.4668000042438507},{"id":"https://openalex.org/C74193536","wikidata":"https://www.wikidata.org/wiki/Q574844","display_name":"Kernel (algebra)","level":2,"score":0.45210000872612},{"id":"https://openalex.org/C206588197","wikidata":"https://www.wikidata.org/wiki/Q846574","display_name":"Reuse","level":2,"score":0.4327999949455261},{"id":"https://openalex.org/C68339613","wikidata":"https://www.wikidata.org/wiki/Q1549489","display_name":"Speedup","level":2,"score":0.3628999888896942},{"id":"https://openalex.org/C127705205","wikidata":"https://www.wikidata.org/wiki/Q5748245","display_name":"Heuristics","level":2,"score":0.3425999879837036},{"id":"https://openalex.org/C97541855","wikidata":"https://www.wikidata.org/wiki/Q830687","display_name":"Reinforcement learning","level":2,"score":0.33180001378059387},{"id":"https://openalex.org/C34388435","wikidata":"https://www.wikidata.org/wiki/Q2267362","display_name":"Bounded function","level":2,"score":0.296099990606308},{"id":"https://openalex.org/C2779960059","wikidata":"https://www.wikidata.org/wiki/Q7113681","display_name":"Overhead (engineering)","level":2,"score":0.29319998621940613},{"id":"https://openalex.org/C80444323","wikidata":"https://www.wikidata.org/wiki/Q2878974","display_name":"Theoretical computer science","level":1,"score":0.2793000042438507},{"id":"https://openalex.org/C2776214188","wikidata":"https://www.wikidata.org/wiki/Q408386","display_name":"Inference","level":2,"score":0.26179999113082886},{"id":"https://openalex.org/C12186640","wikidata":"https://www.wikidata.org/wiki/Q6815743","display_name":"Memory model","level":3,"score":0.2590999901294708},{"id":"https://openalex.org/C2776014549","wikidata":"https://www.wikidata.org/wiki/Q3050847","display_name":"Consolidation (business)","level":2,"score":0.2574000060558319},{"id":"https://openalex.org/C177225278","wikidata":"https://www.wikidata.org/wiki/Q192674","display_name":"Factoring","level":2,"score":0.2565999925136566}],"mesh":[],"locations_count":2,"locations":[{"id":"pmh:oai:arXiv.org:2609.35568","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2609.35568","pdf_url":"https://arxiv.org/pdf/2609.35568","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":"doi:10.48550/arxiv.2609.35568","doi":"https://doi.org/10.48550/arxiv.2609.35568","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2609.35568","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:2609.35568","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2609.35568","pdf_url":"https://arxiv.org/pdf/2609.35568","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":[],"study_designs":[],"awards":[],"funders":[],"has_content":{"grobid_xml":true,"pdf":true},"content_urls":{"pdf":"https://content.openalex.org/works/W7214807286.pdf","grobid_xml":"https://content.openalex.org/works/W7214807286.grobid-xml"},"referenced_works_count":16,"referenced_works":["https://openalex.org/W4407683692","https://openalex.org/W4407831980","https://openalex.org/W4416017756","https://openalex.org/W7077519595","https://openalex.org/W7118792955","https://openalex.org/W7123824377","https://openalex.org/W7125488445","https://openalex.org/W7127140986","https://openalex.org/W7127981264","https://openalex.org/W7133216078","https://openalex.org/W7135099534","https://openalex.org/W7140912173","https://openalex.org/W7160924037","https://openalex.org/W7167229350","https://openalex.org/W7167636339","https://openalex.org/W7171876420"],"related_works":["https://openalex.org/W7223148865","https://openalex.org/W7135099534","https://openalex.org/W4417429088","https://openalex.org/W7167636339","https://openalex.org/W7167229350","https://openalex.org/W7163996892","https://openalex.org/W7223276455","https://openalex.org/W7127140986","https://openalex.org/W7223045346","https://openalex.org/W7171876420"],"abstract_inverted_index":{"High-performance":[0],"kernels":[1],"underpin":[2],"efficient":[3],"accelerator":[4],"execution":[5,35,180],"but":[6,58,75],"require":[7],"expert":[8,62],"tuning":[9],"and":[10,37,64,82,120,158,165,218,225],"lengthy":[11],"manual":[12],"optimization":[13],"cycles.":[14],"LLM":[15],"coding":[16],"agents":[17,69,222],"promise":[18],"automation,":[19],"yet":[20],"their":[21,76],"CUDA":[22],"knowledge":[23,228],"transfers":[24],"poorly":[25],"to":[26,55,163,223],"data-scarce":[27],"domain-specific":[28],"architectures":[29],"(DSAs)":[30],"such":[31],"as":[32],"NPUs,":[33],"whose":[34],"models":[36],"memory":[38],"hierarchies":[39],"differ":[40],"substantially":[41],"from":[42],"those":[43],"of":[44,191],"GPUs.":[45],"To":[46],"address":[47],"this":[48],"transfer":[49],"gap,":[50],"post-training":[51],"methods":[52],"adapt":[53,71],"LLMs":[54],"NPU":[56,133],"programming":[57],"depend":[59],"on":[60,207],"scarce":[61],"data":[63],"substantial":[65],"training":[66],"compute.":[67],"Memory-learning":[68],"instead":[70],"through":[72],"external":[73],"memory,":[74],"uniform":[77],"credit":[78,150,216],"assignment":[79,217],"gives":[80],"adopted":[81],"unused":[83],"experiences":[84,102],"the":[85,185,204],"same":[86],"reward":[87],"target,":[88],"potentially":[89],"biasing":[90],"subsequent":[91],"retrieval":[92,118],"rankings.":[93],"Moreover,":[94],"when":[95],"learned":[96],"values":[97],"guide":[98],"only":[99],"retrieval,":[100],"high-value":[101],"that":[103,214],"generalize":[104],"across":[105,161,229],"operators":[106,162,193],"must":[107],"be":[108],"retrieved":[109],"repeatedly":[110],"rather":[111],"than":[112],"retained":[113],"in":[114],"context,":[115],"thereby":[116],"increasing":[117],"overhead":[119],"weakening":[121],"cross-task":[122],"guidance.":[123],"We":[124],"therefore":[125],"present":[126],"SAGE,":[127],"a":[128,170,178,200],"persistent":[129],"self-improving":[130],"agent":[131],"for":[132,148,184],"kernel":[134,145],"synthesis.":[135],"Adoption-Traced":[136],"Utility":[137],"estimation":[138],"(ATU)":[139],"combines":[140],"explicit":[141],"adoption":[142,160],"records":[143],"with":[144,189],"evaluation":[146],"outcomes":[147],"adoption-aware":[149,215],"assignment.":[151],"Utility-Gated":[152],"Consolidation":[153],"(UGC)":[154],"uses":[155],"positive":[156],"utility":[157],"repeated":[159],"select":[164],"abstract":[166],"reusable":[167],"rules":[168],"into":[169],"bounded":[171],"resident":[172],"context.":[173],"On":[174],"NPUKernelBench,":[175],"SAGE":[176,198],"achieves":[177,199],"95.5%":[179],"rate":[181],"versus":[182],"84.1%":[183],"strongest":[186],"controlled":[187],"baseline,":[188],"86.9%":[190],"solved":[192],"outperforming":[194],"torch_npu.":[195],"With":[196],"GLM-5.3,":[197],"43.99x":[201],"speedup":[202],"over":[203],"torch_npu":[205],"reference":[206],"sparse":[208],"flash":[209],"attention.":[210],"These":[211],"results":[212],"show":[213],"selective":[219],"consolidation":[220],"enable":[221],"accumulate":[224],"reuse":[226],"hardware-specific":[227],"tasks.":[230]},"counts_by_year":[],"updated_date":"2026-11-07T11:25:07.637180","created_date":"2026-09-30T00:00:00"}