{"id":"https://openalex.org/W7127094184","doi":"https://doi.org/10.48550/arxiv.2601.22813","title":"Quartet II: Accurate LLM Pre-Training in NVFP4 by Improved Unbiased Gradient Estimation","display_name":"Quartet II: Accurate LLM Pre-Training in NVFP4 by Improved Unbiased Gradient Estimation","publication_year":2026,"publication_date":"2026-01-30","ids":{"openalex":"https://openalex.org/W7127094184","doi":"https://doi.org/10.48550/arxiv.2601.22813"},"language":null,"primary_location":{"id":"pmh:oai:arXiv.org:2601.22813","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2601.22813","pdf_url":"https://arxiv.org/pdf/2601.22813","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/2601.22813","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5111111342","display_name":"Andrei Panferov","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I157556583","display_name":"Institute of Science and Technology Austria","ror":"https://ror.org/03gnh5541","country_code":"AT","type":"education","lineage":["https://openalex.org/I157556583"]}],"countries":["AT"],"is_corresponding":false,"raw_author_name":"Panferov, Andrei","raw_affiliation_strings":["IST-DASLab/Quartet-II","Institute of Science and Technology Austria"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"IST-DASLab/Quartet-II","institution_ids":[]},{"raw_affiliation_string":"Institute of Science and Technology Austria","institution_ids":["https://openalex.org/I157556583"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5070761496","display_name":"Erik Schultheis","orcid":"https://orcid.org/0000-0003-1685-8397","observed_orcids":["https://orcid.org/0000-0003-1685-8397"]},"institutions":[{"id":"https://openalex.org/I157556583","display_name":"Institute of Science and Technology Austria","ror":"https://ror.org/03gnh5541","country_code":"AT","type":"education","lineage":["https://openalex.org/I157556583"]}],"countries":["AT"],"is_corresponding":false,"raw_author_name":"Schultheis, Erik","raw_affiliation_strings":["Institute of Science and Technology Austria"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Institute of Science and Technology Austria","institution_ids":["https://openalex.org/I157556583"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5002754136","display_name":"Soroush Tabesh","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I157556583","display_name":"Institute of Science and Technology Austria","ror":"https://ror.org/03gnh5541","country_code":"AT","type":"education","lineage":["https://openalex.org/I157556583"]}],"countries":["AT"],"is_corresponding":false,"raw_author_name":"Tabesh, Soroush","raw_affiliation_strings":["Institute of Science and Technology Austria"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Institute of Science and Technology Austria","institution_ids":["https://openalex.org/I157556583"]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5083822059","display_name":"Dan Alistarh","orcid":"https://orcid.org/0000-0003-3650-940X","observed_orcids":["https://orcid.org/0000-0003-3650-940X"]},"institutions":[{"id":"https://openalex.org/I1328219102","display_name":"Red Hat (United States)","ror":"https://ror.org/03ncs3316","country_code":"US","type":"company","lineage":["https://openalex.org/I1328219102"]},{"id":"https://openalex.org/I157556583","display_name":"Institute of Science and Technology Austria","ror":"https://ror.org/03gnh5541","country_code":"AT","type":"education","lineage":["https://openalex.org/I157556583"]}],"countries":["AT","US"],"is_corresponding":false,"raw_author_name":"Alistarh, Dan","raw_affiliation_strings":["Institute of Science and Technology Austria","Red Hat AI"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Institute of Science and Technology Austria","institution_ids":["https://openalex.org/I157556583"]},{"raw_affiliation_string":"Red Hat AI","institution_ids":["https://openalex.org/I1328219102"]}]}],"institutions":[],"countries_distinct_count":2,"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/T11697","display_name":"Numerical Methods and Algorithms","score":0.888700008392334,"subfield":{"id":"https://openalex.org/subfields/1703","display_name":"Computational Theory and Mathematics"},"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/T11697","display_name":"Numerical Methods and Algorithms","score":0.888700008392334,"subfield":{"id":"https://openalex.org/subfields/1703","display_name":"Computational Theory and Mathematics"},"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/T10036","display_name":"Advanced Neural Network Applications","score":0.06539999693632126,"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":0.00989999994635582,"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"}}],"keywords":[{"id":"https://openalex.org/keywords/large-language-models","display_name":"large language models","score":0.8870000243186951},{"id":"https://openalex.org/keywords/unbiased-gradient-estimation","display_name":"unbiased gradient estimation","score":0.9279999732971191},{"id":"https://openalex.org/keywords/stochastic-rounding","display_name":"stochastic rounding","score":0.828000009059906},{"id":"https://openalex.org/keywords/low-precision-arithmetic","display_name":"low-precision arithmetic","score":0.6380000114440918},{"id":"https://openalex.org/keywords/quantization-error","display_name":"quantization error","score":0.6129999756813049}],"concepts":[{"id":"https://openalex.org/C28855332","wikidata":"https://www.wikidata.org/wiki/Q198099","display_name":"Quantization (signal processing)","level":2,"score":0.738099992275238},{"id":"https://openalex.org/C136625980","wikidata":"https://www.wikidata.org/wiki/Q663208","display_name":"Rounding","level":2,"score":0.6980999708175659},{"id":"https://openalex.org/C11413529","wikidata":"https://www.wikidata.org/wiki/Q8366","display_name":"Algorithm","level":1,"score":0.6682000160217285},{"id":"https://openalex.org/C68339613","wikidata":"https://www.wikidata.org/wiki/Q1549489","display_name":"Speedup","level":2,"score":0.6460000276565552},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.5806999802589417},{"id":"https://openalex.org/C2776359362","wikidata":"https://www.wikidata.org/wiki/Q2145286","display_name":"Representation (politics)","level":3,"score":0.3813000023365021},{"id":"https://openalex.org/C2776760102","wikidata":"https://www.wikidata.org/wiki/Q5139990","display_name":"Code (set theory)","level":3,"score":0.3720000088214874},{"id":"https://openalex.org/C33923547","wikidata":"https://www.wikidata.org/wiki/Q395","display_name":"Mathematics","level":0,"score":0.36239999532699585},{"id":"https://openalex.org/C77618280","wikidata":"https://www.wikidata.org/wiki/Q1155772","display_name":"Scheme (mathematics)","level":2,"score":0.3361999988555908},{"id":"https://openalex.org/C61005703","wikidata":"https://www.wikidata.org/wiki/Q2145211","display_name":"Round-off error","level":2,"score":0.3294000029563904},{"id":"https://openalex.org/C43126263","wikidata":"https://www.wikidata.org/wiki/Q128751","display_name":"Source code","level":2,"score":0.30250000953674316},{"id":"https://openalex.org/C106487976","wikidata":"https://www.wikidata.org/wiki/Q685816","display_name":"Matrix (chemical analysis)","level":2,"score":0.2994999885559082},{"id":"https://openalex.org/C103275481","wikidata":"https://www.wikidata.org/wiki/Q6787889","display_name":"Matrix representation","level":3,"score":0.29829999804496765},{"id":"https://openalex.org/C45374587","wikidata":"https://www.wikidata.org/wiki/Q12525525","display_name":"Computation","level":2,"score":0.2955999970436096},{"id":"https://openalex.org/C80444323","wikidata":"https://www.wikidata.org/wiki/Q2878974","display_name":"Theoretical computer science","level":1,"score":0.2888999879360199},{"id":"https://openalex.org/C28826006","wikidata":"https://www.wikidata.org/wiki/Q33521","display_name":"Applied mathematics","level":1,"score":0.2822999954223633},{"id":"https://openalex.org/C191795146","wikidata":"https://www.wikidata.org/wiki/Q3878446","display_name":"Norm (philosophy)","level":2,"score":0.27390000224113464},{"id":"https://openalex.org/C206688291","wikidata":"https://www.wikidata.org/wiki/Q7617819","display_name":"Stochastic gradient descent","level":3,"score":0.26159998774528503},{"id":"https://openalex.org/C51632099","wikidata":"https://www.wikidata.org/wiki/Q3985153","display_name":"Training set","level":2,"score":0.25920000672340393}],"mesh":[],"locations_count":2,"locations":[{"id":"pmh:oai:arXiv.org:2601.22813","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2601.22813","pdf_url":"https://arxiv.org/pdf/2601.22813","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.2601.22813","doi":"https://doi.org/10.48550/arxiv.2601.22813","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2601.22813","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":null,"raw_source_name":null,"raw_type":"Preprint"}],"best_oa_location":{"id":"pmh:oai:arXiv.org:2601.22813","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2601.22813","pdf_url":"https://arxiv.org/pdf/2601.22813","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":[{"id":"https://openalex.org/G4125750076","display_name":null,"funder_award_id":"10.55776/COE12","funder_id":"https://openalex.org/F4320321181","funder_display_name":"Austrian Science Fund"},{"id":"https://openalex.org/G7635477404","display_name":"Bilateral Artificial Intelligence","funder_award_id":"COE12","funder_id":"https://openalex.org/F4320321181","funder_display_name":"Austrian Science Fund"},{"id":"https://openalex.org/G8590477180","display_name":null,"funder_award_id":"10.55776","funder_id":"https://openalex.org/F4320321181","funder_display_name":"Austrian Science Fund"}],"funders":[{"id":"https://openalex.org/F4320309480","display_name":"Nvidia","ror":"https://ror.org/03jdj4y14"},{"id":"https://openalex.org/F4320321181","display_name":"Austrian Science Fund","ror":"https://ror.org/013tf3c58"},{"id":"https://openalex.org/F4320331518","display_name":"Centro Svizzero di Calcolo Scientifico","ror":"https://ror.org/04rzmms09"},{"id":"https://openalex.org/F4320331617","display_name":"National Supercomputing Centre Singapore","ror":"https://ror.org/05qbsk739"}],"has_content":{"grobid_xml":true,"pdf":true},"content_urls":{"pdf":"https://content.openalex.org/works/W7127094184.pdf","grobid_xml":"https://content.openalex.org/works/W7127094184.grobid-xml"},"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"The":[0],"NVFP4":[1,78],"lower-precision":[2],"format,":[3],"supported":[4],"in":[5,42,77],"hardware":[6],"by":[7,51],"NVIDIA":[8,177],"Blackwell":[9,178],"GPUs,":[10],"promises":[11],"to":[12,59,165,182],"allow,":[13],"for":[14,74,85,109,174],"the":[15,36,69,72,133,137],"first":[16],"time,":[17],"end-to-end":[18,160],"fully-quantized":[19],"pre-training":[20],"of":[21,35,39,44,71],"massive":[22],"models":[23],"such":[24],"as":[25],"LLMs.":[26],"Yet,":[27],"existing":[28],"quantized":[29,48,75],"training":[30,76,148,162],"methods":[31],"still":[32],"sacrifice":[33],"some":[34],"representation":[37],"capacity":[38],"this":[40,66],"format":[41],"favor":[43],"more":[45,92],"accurate":[46],"unbiased":[47,82],"gradient":[49,124],"estimation":[50,125],"stochastic":[52],"rounding":[53],"(SR),":[54],"losing":[55],"noticeable":[56],"accuracy":[57],"relative":[58],"standard":[60],"FP16":[61],"and":[62,135],"FP8":[63],"training.":[64],"In":[65,140],"paper,":[67],"improve":[68],"state":[70],"art":[73],"via":[79],"a":[80,104],"novel":[81,105],"quantization":[83,96,107],"routine":[84],"micro-scaled":[86],"formats,":[87],"called":[88,112],"MS-EDEN,":[89],"that":[90,118],"has":[91],"than":[93,98],"2x":[94],"lower":[95],"error":[97],"SR.":[99],"We":[100,115,154,171],"integrate":[101],"it":[102],"into":[103],"fully-NVFP4":[106],"scheme":[108],"linear":[110],"layers,":[111],"Quartet":[113,119,157],"II.":[114],"show":[116],"analytically":[117],"II":[120,158],"achieves":[121],"consistently":[122],"better":[123],"across":[126],"all":[127],"major":[128],"matrix":[129],"multiplications,":[130],"both":[131],"on":[132,136,159,168,176],"forward":[134],"backward":[138],"passes.":[139],"addition,":[141],"our":[142],"proposal":[143],"synergizes":[144],"well":[145],"with":[146,163,180],"recent":[147],"improvements":[149],"aimed":[150],"specifically":[151],"at":[152,191],"NVFP4.":[153],"further":[155],"validate":[156],"LLM":[161],"up":[164,181],"1.9B":[166],"parameters":[167],"38B":[169],"tokens.":[170],"provide":[172],"kernels":[173],"execution":[175],"GPUs":[179],"4.2x":[183],"speedup":[184],"over":[185],"BF16.":[186],"Our":[187],"code":[188],"is":[189],"available":[190],"https://github.com/IST-DASLab/Quartet-II":[192],".":[193]},"counts_by_year":[],"updated_date":"2026-10-15T10:00:46.086208","created_date":"2026-02-03T00:00:00"}