{"id":"https://openalex.org/W4412888196","doi":"https://doi.org/10.18653/v1/2025.findings-acl.776","title":"AMXFP4: Taming Activation Outliers with Asymmetric Microscaling Floating-Point for 4-bit LLM Inference","display_name":"AMXFP4: Taming Activation Outliers with Asymmetric Microscaling Floating-Point for 4-bit LLM Inference","publication_year":2025,"publication_date":"2025-01-01","ids":{"openalex":"https://openalex.org/W4412888196","doi":"https://doi.org/10.18653/v1/2025.findings-acl.776"},"language":"en","primary_location":{"id":"doi:10.18653/v1/2025.findings-acl.776","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2025.findings-acl.776","pdf_url":"https://aclanthology.org/2025.findings-acl.776.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":"Findings of the Association for Computational Linguistics: ACL 2025","raw_type":"proceedings-article"},"type":"conference-paper","indexed_in":["crossref"],"open_access":{"is_oa":true,"oa_status":"gold","oa_url":"https://aclanthology.org/2025.findings-acl.776.pdf","any_repository_has_fulltext":null},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5036363860","display_name":"Janghwan Lee","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Janghwan Lee","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5101918926","display_name":"Jiwoong Park","orcid":"https://orcid.org/0000-0002-0511-1442"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Jiwoong Park","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5100757430","display_name":"Jin-Seok Kim","orcid":"https://orcid.org/0000-0001-5247-1603"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Jinseok Kim","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":null,"display_name":"Yongjik Kim","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Yongjik Kim","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5110546049","display_name":"Jungju Oh","orcid":"https://orcid.org/0009-0009-9451-251X"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Jungju Oh","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5089478486","display_name":"Jinwook Oh","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Jinwook Oh","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5078440061","display_name":"Jungwook Choi","orcid":"https://orcid.org/0000-0002-3075-8694"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Jungwook Choi","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":true,"cited_by_count":1,"citation_normalized_percentile":null,"cited_by_percentile_year":null,"biblio":{"volume":null,"issue":null,"first_page":"14993","last_page":"15013"},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T10558","display_name":"Advancements in Semiconductor Devices and Circuit Design","score":0.9764999747276306,"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/T10558","display_name":"Advancements in Semiconductor Devices and Circuit Design","score":0.9764999747276306,"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/T10472","display_name":"Semiconductor materials and devices","score":0.949400007724762,"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/T11206","display_name":"Model Reduction and Neural Networks","score":0.9171000123023987,"subfield":{"id":"https://openalex.org/subfields/3109","display_name":"Statistical and Nonlinear Physics"},"field":{"id":"https://openalex.org/fields/31","display_name":"Physics and Astronomy"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}}],"keywords":[{"id":"https://openalex.org/keywords/inference","display_name":"Inference","score":0.7594294548034668},{"id":"https://openalex.org/keywords/outlier","display_name":"Outlier","score":0.6806095838546753},{"id":"https://openalex.org/keywords/computer-science","display_name":"Computer science","score":0.6318339109420776},{"id":"https://openalex.org/keywords/bit","display_name":"Bit (key)","score":0.5357519388198853},{"id":"https://openalex.org/keywords/point","display_name":"Point (geometry)","score":0.47651708126068115},{"id":"https://openalex.org/keywords/algorithm","display_name":"Algorithm","score":0.3465626835823059},{"id":"https://openalex.org/keywords/artificial-intelligence","display_name":"Artificial intelligence","score":0.31689196825027466},{"id":"https://openalex.org/keywords/mathematics","display_name":"Mathematics","score":0.18332546949386597},{"id":"https://openalex.org/keywords/computer-network","display_name":"Computer network","score":0.11090633273124695}],"concepts":[{"id":"https://openalex.org/C2776214188","wikidata":"https://www.wikidata.org/wiki/Q408386","display_name":"Inference","level":2,"score":0.7594294548034668},{"id":"https://openalex.org/C79337645","wikidata":"https://www.wikidata.org/wiki/Q779824","display_name":"Outlier","level":2,"score":0.6806095838546753},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.6318339109420776},{"id":"https://openalex.org/C117011727","wikidata":"https://www.wikidata.org/wiki/Q1278488","display_name":"Bit (key)","level":2,"score":0.5357519388198853},{"id":"https://openalex.org/C28719098","wikidata":"https://www.wikidata.org/wiki/Q44946","display_name":"Point (geometry)","level":2,"score":0.47651708126068115},{"id":"https://openalex.org/C11413529","wikidata":"https://www.wikidata.org/wiki/Q8366","display_name":"Algorithm","level":1,"score":0.3465626835823059},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.31689196825027466},{"id":"https://openalex.org/C33923547","wikidata":"https://www.wikidata.org/wiki/Q395","display_name":"Mathematics","level":0,"score":0.18332546949386597},{"id":"https://openalex.org/C31258907","wikidata":"https://www.wikidata.org/wiki/Q1301371","display_name":"Computer network","level":1,"score":0.11090633273124695},{"id":"https://openalex.org/C2524010","wikidata":"https://www.wikidata.org/wiki/Q8087","display_name":"Geometry","level":1,"score":0.0}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.18653/v1/2025.findings-acl.776","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2025.findings-acl.776","pdf_url":"https://aclanthology.org/2025.findings-acl.776.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":"Findings of the Association for Computational Linguistics: ACL 2025","raw_type":"proceedings-article"}],"best_oa_location":{"id":"doi:10.18653/v1/2025.findings-acl.776","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2025.findings-acl.776","pdf_url":"https://aclanthology.org/2025.findings-acl.776.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":"Findings of the Association for Computational Linguistics: ACL 2025","raw_type":"proceedings-article"},"sustainable_development_goals":[],"awards":[],"funders":[],"has_content":{"grobid_xml":true,"pdf":true},"content_urls":{"pdf":"https://content.openalex.org/works/W4412888196.pdf","grobid_xml":"https://content.openalex.org/works/W4412888196.grobid-xml"},"referenced_works_count":0,"referenced_works":[],"related_works":["https://openalex.org/W4327546585","https://openalex.org/W2051487156","https://openalex.org/W2411923897","https://openalex.org/W4394546135","https://openalex.org/W4285347720","https://openalex.org/W4200259850","https://openalex.org/W2333831899","https://openalex.org/W2073681303","https://openalex.org/W2484894494","https://openalex.org/W2367385042"],"abstract_inverted_index":{"As":[0],"large":[1],"language":[2],"models":[3],"(LLMs)":[4],"grow":[5],"in":[6,50],"parameter":[7],"size":[8],"and":[9,46,83,157],"context":[10],"length,":[11],"computation":[12],"precision":[13],"has":[14],"been":[15],"reduced":[16],"from":[17],"16-bit":[18],"to":[19,21,31],"4bit":[20],"improve":[22],"inference":[23],"efficiency.However,":[24],"this":[25,37,92],"reduction":[26],"causes":[27],"accuracy":[28],"degradation":[29],"due":[30],"activation":[32,107],"outliers.Rotation-based":[33],"INT4":[34],"methods":[35,160],"address":[36,119],"via":[38],"matrix":[39,67],"calibration,":[40],"but":[41],"they":[42],"introduce":[43],"multi-hour":[44],"overheads":[45],"leave":[47],"key":[48],"computations":[49],"full":[51],"precision.Microscaling":[52],"(MX)":[53],"floating-point":[54],"(FP)":[55],"formats":[56,88],"offer":[57],"fine-grained":[58],"representation":[59],"with":[60],"a":[61,124],"shared":[62,135],"scale,":[63],"enabling":[64],"fully":[65],"quantized":[66],"multiplications":[68],"through":[69],"direct":[70],"casting":[71],"without":[72,137],"calibration.However,":[73],"existing":[74],"research":[75],"shows":[76],"unsatisfactory":[77],"empirical":[78],"results":[79],"for":[80],"MXFP4":[81,152,170],"inference,":[82],"the":[84,96,100,113],"robustness":[85],"of":[86,99,115],"MX":[87,101],"remains":[89],"largely":[90],"unexplored.In":[91],"work,":[93],"we":[94,121],"uncover":[95],"fundamental":[97],"tradeoffs":[98],"format:":[102],"while":[103,147],"it":[104,109],"effectively":[105],"suppresses":[106],"outliers,":[108],"does":[110],"so":[111],"at":[112],"cost":[114,146],"increased":[116],"group-wise":[117],"asymmetry.To":[118],"this,":[120],"propose":[122],"AMXFP4,":[123],"4-bit":[125],"asymmetric":[126,134],"FP":[127],"format":[128],"that":[129],"handles":[130],"both":[131],"issues":[132],"using":[133],"scales,":[136],"requiring":[138],"calibration.Our":[139],"custom":[140],"MAC":[141],"engine":[142],"adds":[143],"negligible":[144],"hardware":[145],"improving":[148],"accuracy:":[149],"AMXFP4":[150],"outperforms":[151],"by":[153,161],"3%":[154],"on":[155,163],"VQA":[156],"exceeds":[158],"rotation-based":[159],"1.6%":[162],"CSQA.It":[164],"also":[165],"surpasses":[166],"recently":[167],"deployed":[168],"commercial":[169],"variants.":[171]},"counts_by_year":[{"year":2026,"cited_by_count":1}],"updated_date":"2026-07-14T23:27:15.235271","created_date":"2025-10-10T00:00:00"}