{"id":"https://openalex.org/W7163911910","doi":"https://doi.org/10.48550/arxiv.2606.07116","title":"OffQ: Taming Structured Outliers in LLM Quantization by Offsetting","display_name":"OffQ: Taming Structured Outliers in LLM Quantization by Offsetting","publication_year":2026,"publication_date":"2026-06-05","ids":{"openalex":"https://openalex.org/W7163911910","doi":"https://doi.org/10.48550/arxiv.2606.07116"},"language":null,"primary_location":{"id":"doi:10.48550/arxiv.2606.07116","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2606.07116","pdf_url":null,"source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":null,"issn":null,"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":"cc-by","license_id":"https://openalex.org/licenses/cc-by","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.2606.07116","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5138200697","display_name":"Haoqi Wang","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Wang, Haoqi","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5138166852","display_name":"Lorenz K. Mueller","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Mueller, Lorenz K.","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5138113255","display_name":"Jiawei Zhuang","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Zhuang, Jiawei","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5049300388","display_name":"Mathieu Salzmann","orcid":"https://orcid.org/0000-0002-8347-8637"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Salzmann, Mathieu","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5138152093","display_name":"Lukas Cavigelli","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Cavigelli, Lukas","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":0.3930000066757202,"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":0.3930000066757202,"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/T11307","display_name":"Domain Adaptation and Few-Shot Learning","score":0.11159999668598175,"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"}},{"id":"https://openalex.org/T10201","display_name":"Speech Recognition and Synthesis","score":0.06689999997615814,"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/outlier","display_name":"Outlier","score":0.8307999968528748},{"id":"https://openalex.org/keywords/quantization","display_name":"Quantization (signal processing)","score":0.7766000032424927},{"id":"https://openalex.org/keywords/inference","display_name":"Inference","score":0.5680000185966492},{"id":"https://openalex.org/keywords/subspace-topology","display_name":"Subspace topology","score":0.4677000045776367},{"id":"https://openalex.org/keywords/anomaly-detection","display_name":"Anomaly detection","score":0.397599995136261},{"id":"https://openalex.org/keywords/vector-quantization","display_name":"Vector quantization","score":0.3698999881744385}],"concepts":[{"id":"https://openalex.org/C79337645","wikidata":"https://www.wikidata.org/wiki/Q779824","display_name":"Outlier","level":2,"score":0.8307999968528748},{"id":"https://openalex.org/C28855332","wikidata":"https://www.wikidata.org/wiki/Q198099","display_name":"Quantization (signal processing)","level":2,"score":0.7766000032424927},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.656000018119812},{"id":"https://openalex.org/C2776214188","wikidata":"https://www.wikidata.org/wiki/Q408386","display_name":"Inference","level":2,"score":0.5680000185966492},{"id":"https://openalex.org/C32834561","wikidata":"https://www.wikidata.org/wiki/Q660730","display_name":"Subspace topology","level":2,"score":0.4677000045776367},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.4187999963760376},{"id":"https://openalex.org/C739882","wikidata":"https://www.wikidata.org/wiki/Q3560506","display_name":"Anomaly detection","level":2,"score":0.397599995136261},{"id":"https://openalex.org/C199833920","wikidata":"https://www.wikidata.org/wiki/Q612536","display_name":"Vector quantization","level":2,"score":0.3698999881744385},{"id":"https://openalex.org/C46686674","wikidata":"https://www.wikidata.org/wiki/Q466303","display_name":"Boosting (machine learning)","level":2,"score":0.36149999499320984},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.35690000653266907},{"id":"https://openalex.org/C11413529","wikidata":"https://www.wikidata.org/wiki/Q8366","display_name":"Algorithm","level":1,"score":0.3538999855518341},{"id":"https://openalex.org/C127162648","wikidata":"https://www.wikidata.org/wiki/Q16858953","display_name":"Channel (broadcasting)","level":2,"score":0.3472999930381775},{"id":"https://openalex.org/C124101348","wikidata":"https://www.wikidata.org/wiki/Q172491","display_name":"Data mining","level":1,"score":0.3221000134944916},{"id":"https://openalex.org/C165696696","wikidata":"https://www.wikidata.org/wiki/Q11287","display_name":"Exploit","level":2,"score":0.30250000953674316}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.48550/arxiv.2606.07116","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2606.07116","pdf_url":null,"source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":null,"issn":null,"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":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":null,"is_accepted":false,"is_published":null,"raw_source_name":null,"raw_type":"Preprint"}],"best_oa_location":{"id":"doi:10.48550/arxiv.2606.07116","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2606.07116","pdf_url":null,"source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":null,"issn":null,"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":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":null,"is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"Preprint"},"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":{"Low-bit":[0],"quantization":[1,54,115],"has":[2],"been":[3],"widely":[4],"adopted":[5],"to":[6,30,35,48],"accelerate":[7],"the":[8,69,103,107],"inference":[9],"of":[10,106,116],"large":[11],"language":[12],"models":[13],"(LLMs)":[14],"by":[15,93],"significantly":[16],"reducing":[17,102],"computational":[18],"cost":[19],"and":[20,76,121,130],"memory":[21],"usage.":[22],"However,":[23],"activation":[24,50],"outliers":[25,51],"pose":[26],"a":[27,45,56,64,72,98],"major":[28],"challenge":[29],"effective":[31,113],"quantization,":[32],"often":[33],"leading":[34],"notable":[36],"performance":[37],"degradation.":[38],"In":[39],"this":[40,89],"paper,":[41],"we":[42],"introduce":[43],"OffQ,":[44],"method":[46],"designed":[47],"mitigate":[49],"in":[52,68],"low-bit":[53,144],"through":[55],"novel":[57],"offsetting":[58,110],"mechanism.":[59],"Specifically,":[60],"OffQ":[61,86,134],"first":[62],"identifies":[63],"low-dimensional":[65],"outlier":[66,91],"subspace":[67],"activations":[70,80],"using":[71,118],"proposed":[73],"top-1":[74],"PCA,":[75],"then":[77,87],"concentrates":[78],"high-magnitude":[79],"into":[81,97],"1":[82],"channel":[83,92],"via":[84],"rotation.":[85],"absorbs":[88],"concentrated":[90],"converting":[94],"its":[95],"magnitude":[96],"shared":[99],"offset,":[100],"thereby":[101],"standard":[104],"deviation":[105],"activations.":[108],"This":[109],"strategy":[111],"enables":[112],"W4A4KV4":[114],"LLMs":[117],"deployment-friendly":[119],"uniform-grid":[120],"uniform-precision":[122],"quantization.":[123],"Extensive":[124],"experiments":[125],"across":[126],"diverse":[127],"LLM":[128],"architectures":[129],"benchmarks":[131],"demonstrate":[132],"that":[133],"outperforms":[135],"state-of-the-art":[136],"baselines,":[137],"consistently":[138],"improving":[139],"model":[140],"accuracy":[141],"while":[142],"preserving":[143],"efficiency.":[145]},"counts_by_year":[],"updated_date":"2026-07-01T06:00:48.157686","created_date":"2026-06-09T00:00:00"}