{"id":"https://openalex.org/W7167636339","doi":"https://doi.org/10.48550/arxiv.2607.04395","title":"NKI-Agent: Domain-Specific Fine-Tuning and Agentic Tool Use for Neuron Kernel Generation","display_name":"NKI-Agent: Domain-Specific Fine-Tuning and Agentic Tool Use for Neuron Kernel Generation","publication_year":2026,"publication_date":"2026-07-05","ids":{"openalex":"https://openalex.org/W7167636339","doi":"https://doi.org/10.48550/arxiv.2607.04395"},"language":null,"primary_location":{"id":"doi:10.48550/arxiv.2607.04395","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2607.04395","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,"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"},"type":"preprint","indexed_in":["datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://doi.org/10.48550/arxiv.2607.04395","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5140208153","display_name":"Junjie Tang","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Tang, Junjie","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5140211077","display_name":"Jun Huan","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Huan, Jun","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5140212685","display_name":"Hao Zhou","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Zhou, Hao","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5140204805","display_name":"Yuhao Zhang","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Zhang, Yuhao","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5140170572","display_name":"Lin Wang","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Wang, Lin","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/T10036","display_name":"Advanced Neural Network Applications","score":0.7296000123023987,"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"}},"topics":[{"id":"https://openalex.org/T10036","display_name":"Advanced Neural Network Applications","score":0.7296000123023987,"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"}},{"id":"https://openalex.org/T10054","display_name":"Parallel Computing and Optimization Techniques","score":0.08590000122785568,"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/T10502","display_name":"Advanced Memory and Neural Computing","score":0.021400000900030136,"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/kernel","display_name":"Kernel (algebra)","score":0.7365000247955322},{"id":"https://openalex.org/keywords/construct","display_name":"Construct (python library)","score":0.5001999735832214},{"id":"https://openalex.org/keywords/benchmark","display_name":"Benchmark (surveying)","score":0.4878999888896942},{"id":"https://openalex.org/keywords/tree-kernel","display_name":"Tree kernel","score":0.435699999332428},{"id":"https://openalex.org/keywords/kernel-method","display_name":"Kernel method","score":0.3345000147819519},{"id":"https://openalex.org/keywords/memory-model","display_name":"Memory model","score":0.30300000309944153}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.7670999765396118},{"id":"https://openalex.org/C74193536","wikidata":"https://www.wikidata.org/wiki/Q574844","display_name":"Kernel (algebra)","level":2,"score":0.7365000247955322},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.5414000153541565},{"id":"https://openalex.org/C2780801425","wikidata":"https://www.wikidata.org/wiki/Q5164392","display_name":"Construct (python library)","level":2,"score":0.5001999735832214},{"id":"https://openalex.org/C185798385","wikidata":"https://www.wikidata.org/wiki/Q1161707","display_name":"Benchmark (surveying)","level":2,"score":0.4878999888896942},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.4399000108242035},{"id":"https://openalex.org/C140417398","wikidata":"https://www.wikidata.org/wiki/Q16933942","display_name":"Tree kernel","level":5,"score":0.435699999332428},{"id":"https://openalex.org/C122280245","wikidata":"https://www.wikidata.org/wiki/Q620622","display_name":"Kernel method","level":3,"score":0.3345000147819519},{"id":"https://openalex.org/C12186640","wikidata":"https://www.wikidata.org/wiki/Q6815743","display_name":"Memory model","level":3,"score":0.30300000309944153},{"id":"https://openalex.org/C120524526","wikidata":"https://www.wikidata.org/wiki/Q1709148","display_name":"Reboot","level":2,"score":0.29760000109672546},{"id":"https://openalex.org/C4679612","wikidata":"https://www.wikidata.org/wiki/Q866298","display_name":"Aggregate (composite)","level":2,"score":0.2786000072956085},{"id":"https://openalex.org/C48677424","wikidata":"https://www.wikidata.org/wiki/Q6888088","display_name":"Mode (computer interface)","level":2,"score":0.26739999651908875},{"id":"https://openalex.org/C113843644","wikidata":"https://www.wikidata.org/wiki/Q901882","display_name":"Interface (matter)","level":4,"score":0.2596000134944916},{"id":"https://openalex.org/C48372109","wikidata":"https://www.wikidata.org/wiki/Q3913","display_name":"Binary number","level":2,"score":0.25369998812675476},{"id":"https://openalex.org/C51632099","wikidata":"https://www.wikidata.org/wiki/Q3985153","display_name":"Training set","level":2,"score":0.25040000677108765}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.48550/arxiv.2607.04395","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2607.04395","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,"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":null,"raw_source_name":null,"raw_type":"Preprint"}],"best_oa_location":{"id":"doi:10.48550/arxiv.2607.04395","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2607.04395","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,"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"},"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":{"Recent":[0],"agentic":[1],"approaches":[2],"to":[3,100,199],"LLM-based":[4],"kernel":[5,24,92,106,210],"generation":[6,25,107],"have":[7],"achieved":[8],"impressive":[9],"results":[10],"on":[11,122,141,205],"CUDA.":[12],"For":[13],"emerging":[14],"AI":[15],"accelerators":[16],"such":[17],"as":[18],"AWS":[19],"Trainium":[20],"and":[21,26,53,111,131],"Inferentia,":[22],"automated":[23],"optimization":[27],"remain":[28],"largely":[29],"unaddressed.":[30],"Writing":[31],"kernels":[32],"for":[33,71,90,109,208],"these":[34],"chips":[35],"via":[36],"the":[37,77,96,142,178],"Neuron":[38,101],"Kernel":[39],"Interface":[40],"(NKI)":[41],"is":[42,150],"particularly":[43],"challenging:":[44],"developers":[45],"must":[46],"navigate":[47],"a":[48,86,114,132,137,163],"multi-engine":[49],"architecture,":[50],"tile-based":[51],"programming,":[52],"explicit":[54],"data":[55],"movement":[56],"across":[57,117],"multi-level":[58],"memory":[59],"hierarchy.":[60],"Moreover,":[61],"no":[62],"publicly-available":[63],"training":[64],"data,":[65],"benchmarks,":[66],"or":[67],"tool-augmented":[68],"agents":[69],"exist":[70],"this":[72],"domain.":[73],"We":[74,94,145,185],"introduce":[75],"NKI-Agent,":[76],"first":[78],"system":[79,134],"combining":[80],"domain-specific":[81],"supervised":[82],"fine-tuning":[83],"(SFT)":[84],"with":[85,127,194],"compile-verify-fix":[87],"agent":[88,160],"loop":[89],"NKI":[91,105],"generation.":[93,211],"adapt":[95],"existing":[97],"CUDA-Agent":[98],"framework":[99],"hardware,":[102,125],"curate":[103],"6,000":[104],"tasks":[108],"training,":[110],"construct":[112],"NKIBench,":[113],"250-task":[115],"benchmark":[116],"three":[118],"difficulty":[119],"levels.":[120],"Evaluated":[121],"real":[123],"Trn1":[124],"NKI-Agent":[126],"Claude":[128,181],"Opus":[129,152],"4.8":[130,153],"rank-aware":[133],"prompt":[135],"achieves":[136,172],"77.3%":[138],"pass":[139,174],"rate":[140,175],"150-task":[143],"NKIBench.":[144],"show":[146,167],"that":[147,168,188],"tool":[148],"use":[149],"critical:":[151],"scores":[154],"6%":[155],"in":[156],"single-shot":[157],"mode":[158],"without":[159],"tools.":[161],"On":[162],"60-task":[164],"subset,":[165],"we":[166],"an":[169],"SFT-trained":[170],"Qwen3-Coder-30B-A3B":[171],"25.0%":[173],"at":[176],"1/100th":[177],"cost,":[179],"outperforming":[180],"Sonnet":[182],"4":[183],"(15.0%).":[184],"also":[186],"report":[187],"Group":[189],"Relative":[190],"Policy":[191],"Optimization":[192],"(GRPO)":[193],"binary":[195],"compilation":[196],"reward":[197,206],"fails":[198],"improve":[200],"over":[201],"SFT,":[202],"providing":[203],"guidance":[204],"design":[207],"RL-based":[209]},"counts_by_year":[],"updated_date":"2026-07-28T07:46:37.118299","created_date":"2026-07-08T00:00:00"}