{"id":"https://openalex.org/W4414359705","doi":"https://doi.org/10.24963/ijcai.2025/920","title":"MPPQ: Enhancing Post-Training Quantization for LLMs via Mixed Supervision, Proxy Rounding, and Pre-Searching","display_name":"MPPQ: Enhancing Post-Training Quantization for LLMs via Mixed Supervision, Proxy Rounding, and Pre-Searching","publication_year":2025,"publication_date":"2025-09-01","ids":{"openalex":"https://openalex.org/W4414359705","doi":"https://doi.org/10.24963/ijcai.2025/920"},"language":"en","primary_location":{"id":"doi:10.24963/ijcai.2025/920","is_oa":false,"landing_page_url":"https://doi.org/10.24963/ijcai.2025/920","pdf_url":null,"source":null,"license":null,"license_id":null,"version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Proceedings of the Thirty-Fourth International Joint Conference on Artificial Intelligence","raw_type":"proceedings-article"},"type":"conference-paper","indexed_in":["crossref"],"open_access":{"is_oa":false,"oa_status":"closed","oa_url":null,"any_repository_has_fulltext":false},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5005577580","display_name":"Mingfei Wei","orcid":"https://orcid.org/0000-0002-7148-2532"},"institutions":[{"id":"https://openalex.org/I21193070","display_name":"Beijing Jiaotong University","ror":"https://ror.org/01yj56c84","country_code":"CN","type":"education","lineage":["https://openalex.org/I21193070"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Mingrun Wei","raw_affiliation_strings":["School of Computer Science and Technology, Beijing Jiaotong University, Beijing, China"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"School of Computer Science and Technology, Beijing Jiaotong University, Beijing, China","institution_ids":["https://openalex.org/I21193070"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5017288103","display_name":"Yeyu Yan","orcid":"https://orcid.org/0000-0002-6288-453X"},"institutions":[{"id":"https://openalex.org/I21193070","display_name":"Beijing Jiaotong University","ror":"https://ror.org/01yj56c84","country_code":"CN","type":"education","lineage":["https://openalex.org/I21193070"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Yeyu Yan","raw_affiliation_strings":["School of Computer Science and Technology, Beijing Jiaotong University, Beijing, China"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"School of Computer Science and Technology, Beijing Jiaotong University, Beijing, China","institution_ids":["https://openalex.org/I21193070"]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5100391418","display_name":"Dong Wang","orcid":"https://orcid.org/0000-0001-7017-8259"},"institutions":[{"id":"https://openalex.org/I21193070","display_name":"Beijing Jiaotong University","ror":"https://ror.org/01yj56c84","country_code":"CN","type":"education","lineage":["https://openalex.org/I21193070"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Dong Wang","raw_affiliation_strings":["School of Computer Science and Technology, Beijing Jiaotong University, Beijing, China"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"School of Computer Science and Technology, Beijing Jiaotong University, Beijing, China","institution_ids":["https://openalex.org/I21193070"]}]}],"institutions":[],"countries_distinct_count":1,"institutions_distinct_count":1,"corresponding_author_ids":[],"corresponding_institution_ids":["https://openalex.org/I21193070"],"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":"8277","last_page":"8285"},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T10181","display_name":"Natural Language Processing Techniques","score":0.9761000275611877,"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/T10181","display_name":"Natural Language Processing Techniques","score":0.9761000275611877,"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/quantization","display_name":"Quantization (signal processing)","score":0.7889999747276306},{"id":"https://openalex.org/keywords/rounding","display_name":"Rounding","score":0.6049000024795532},{"id":"https://openalex.org/keywords/initialization","display_name":"Initialization","score":0.4787999987602234},{"id":"https://openalex.org/keywords/scaling","display_name":"Scaling","score":0.4090999960899353},{"id":"https://openalex.org/keywords/clipping","display_name":"Clipping (morphology)","score":0.3991999924182892},{"id":"https://openalex.org/keywords/regularization","display_name":"Regularization (linguistics)","score":0.3515999913215637}],"concepts":[{"id":"https://openalex.org/C28855332","wikidata":"https://www.wikidata.org/wiki/Q198099","display_name":"Quantization (signal processing)","level":2,"score":0.7889999747276306},{"id":"https://openalex.org/C136625980","wikidata":"https://www.wikidata.org/wiki/Q663208","display_name":"Rounding","level":2,"score":0.6049000024795532},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.48030000925064087},{"id":"https://openalex.org/C114466953","wikidata":"https://www.wikidata.org/wiki/Q6034165","display_name":"Initialization","level":2,"score":0.4787999987602234},{"id":"https://openalex.org/C11413529","wikidata":"https://www.wikidata.org/wiki/Q8366","display_name":"Algorithm","level":1,"score":0.46639999747276306},{"id":"https://openalex.org/C33923547","wikidata":"https://www.wikidata.org/wiki/Q395","display_name":"Mathematics","level":0,"score":0.4104999899864197},{"id":"https://openalex.org/C99844830","wikidata":"https://www.wikidata.org/wiki/Q102441924","display_name":"Scaling","level":2,"score":0.4090999960899353},{"id":"https://openalex.org/C2776848632","wikidata":"https://www.wikidata.org/wiki/Q853463","display_name":"Clipping (morphology)","level":2,"score":0.3991999924182892},{"id":"https://openalex.org/C2776135515","wikidata":"https://www.wikidata.org/wiki/Q17143721","display_name":"Regularization (linguistics)","level":2,"score":0.3515999913215637},{"id":"https://openalex.org/C126255220","wikidata":"https://www.wikidata.org/wiki/Q141495","display_name":"Mathematical optimization","level":1,"score":0.3400999903678894},{"id":"https://openalex.org/C93372532","wikidata":"https://www.wikidata.org/wiki/Q6552455","display_name":"Linde\u2013Buzo\u2013Gray algorithm","level":3,"score":0.329800009727478},{"id":"https://openalex.org/C199833920","wikidata":"https://www.wikidata.org/wiki/Q612536","display_name":"Vector quantization","level":2,"score":0.32839998602867126},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.27730000019073486},{"id":"https://openalex.org/C137293760","wikidata":"https://www.wikidata.org/wiki/Q3621696","display_name":"Language model","level":2,"score":0.25589999556541443}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.24963/ijcai.2025/920","is_oa":false,"landing_page_url":"https://doi.org/10.24963/ijcai.2025/920","pdf_url":null,"source":null,"license":null,"license_id":null,"version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Proceedings of the Thirty-Fourth International Joint Conference on Artificial Intelligence","raw_type":"proceedings-article"}],"best_oa_location":null,"sustainable_development_goals":[],"awards":[],"funders":[],"has_content":{"pdf":false,"grobid_xml":false},"content_urls":null,"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Recently,":[0],"post-training":[1],"quantization":[2,74,103,121,145],"(PTQ)":[3],"methods":[4,142],"for":[5,79,168],"large":[6],"language":[7],"models":[8],"(LLMs)":[9],"primarily":[10],"focus":[11],"on":[12,63],"tackling":[13],"the":[14,30,69,95,101,149,164,169],"challenges":[15],"caused":[16,72],"by":[17,73],"outliers.":[18],"Scaling":[19],"transformation":[20],"has":[21],"proven":[22],"to":[23,28,67,99,116,157],"be":[24,154],"effective":[25],"while":[26,75,125],"how":[27],"enhance":[29],"performance":[31],"of":[32,129,151,163],"extremely":[33],"low-bitwidth":[34],"(e.g.,":[35],"2-bit)":[36],"PTQ":[37,48],"under":[38],"it":[39],"remains":[40],"largely":[41],"unexplored.":[42],"In":[43],"this":[44],"work,":[45],"a":[46,85,109],"new":[47],"framework,":[49],"namely":[50],"MPPQ,":[51],"is":[52,114,173],"established.":[53],"Specifically,":[54],"MPPQ":[55,138],"first":[56],"proposes":[57],"an":[58],"enhanced":[59],"reconstruction":[60],"loss":[61],"based":[62],"Mixed":[64],"metric":[65],"supervision":[66],"mitigate":[68],"distribution":[70],"inconsistency":[71],"providing":[76],"strong":[77],"regularization":[78],"learnable":[80],"parameters.":[81],"Secondly,":[82],"we":[83],"introduce":[84],"Proxy-based":[86],"adaptive":[87],"rounding":[88],"scheme":[89],"in":[90,143],"weight":[91],"quantization,":[92],"which":[93,172],"replaces":[94],"round-to-nearest":[96],"(RTN)":[97],"function":[98],"minimize":[100],"overall":[102],"errors":[104],"through":[105],"element-wise":[106],"scaling.":[107],"Furthermore,":[108],"factor":[110],"coarse":[111],"Pre-searching":[112],"mechanism":[113],"presented":[115],"ensure":[117],"proper":[118],"coordination":[119],"between":[120],"and":[122],"clipping":[123,130],"patterns,":[124],"achieving":[126],"optimal":[127],"initialization":[128],"factors":[131],"before":[132],"training.":[133],"Extensive":[134],"experiments":[135],"show":[136],"that":[137],"consistently":[139],"outperforms":[140],"state-of-the-art":[141],"low-bit":[144],"settings.":[146],"For":[147],"instance,":[148],"perplexity":[150],"WikiText2":[152],"can":[153],"dramatically":[155],"reduced":[156],"8.85":[158],"(3.9":[159],"\u2193":[160],"vs":[161],"12.75":[162],"latest":[165],"method,":[166],"LRQuant)":[167],"LLaMA-2-7B":[170],"model,":[171],"quantized":[174],"with":[175],"W4A4.":[176]},"counts_by_year":[],"updated_date":"2026-07-14T23:27:15.235271","created_date":"2025-10-10T00:00:00"}