{"id":"https://openalex.org/W7204589553","doi":"https://doi.org/10.48550/arxiv.2608.26581","title":"Activation Outliers Matter: Robust Recovery for Quantized Multimodal LLMs","display_name":"Activation Outliers Matter: Robust Recovery for Quantized Multimodal LLMs","publication_year":2026,"publication_date":"2026-08-27","ids":{"openalex":"https://openalex.org/W7204589553","doi":"https://doi.org/10.48550/arxiv.2608.26581"},"language":"en","primary_location":{"id":"pmh:oai:arXiv.org:2608.26581","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2608.26581","pdf_url":"https://arxiv.org/pdf/2608.26581","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":"public-domain","license_id":"https://openalex.org/licenses/public-domain","version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},"volume":null,"type":"preprint","is_traditional":true,"indexed_in":["arxiv","datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://arxiv.org/pdf/2608.26581","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5003406651","display_name":"Tanzila Rahman","orcid":"https://orcid.org/0000-0001-8763-8805","observed_orcids":["https://orcid.org/0000-0001-8763-8805"]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Rahman, Tanzila","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5083845399","display_name":"Mehran Taghian","orcid":"https://orcid.org/0000-0003-4996-353X","observed_orcids":["https://orcid.org/0000-0003-4996-353X"]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Jazi, Mehran Taghian","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":null,"display_name":"Peng, Yunke","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Peng, Yunke","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5100297573","display_name":"Zhuang Ma","orcid":"https://orcid.org/0009-0002-6166-8998","observed_orcids":["https://orcid.org/0009-0002-6166-8998"]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Ma, Zhuang","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5022932578","display_name":"Anandharaju Durai Raju","orcid":"https://orcid.org/0000-0001-9873-5870","observed_orcids":["https://orcid.org/0000-0001-9873-5870"]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Raju, Anandharaju Durai","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5148475113","display_name":"Yao Wang","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Wang, Yao","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5143924857","display_name":"Xing Huang","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Huang, Xing","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5085639863","display_name":"Hei Yi Mak","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Mak, Hei Yi","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5003674260","display_name":"Shadan Golestan","orcid":"https://orcid.org/0000-0001-7906-2287","observed_orcids":["https://orcid.org/0000-0001-7906-2287"]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Golestan, Shadan","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5017657388","display_name":"Hoang Le","orcid":"https://orcid.org/0000-0003-4785-2529","observed_orcids":["https://orcid.org/0000-0003-4785-2529"]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Le, Hoang","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5051789667","display_name":"Yonghan Dong","orcid":"https://orcid.org/0000-0003-0866-0663","observed_orcids":["https://orcid.org/0000-0003-0866-0663"]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Dong, Yonghan","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5148568838","display_name":"Wei Guo","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Guo, Wei","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5130965301","display_name":"Yaoyuan Wang","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Wang, Yaoyuan","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":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":0.975600004196167,"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":0.975600004196167,"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/T10363","display_name":"Low-power high-performance VLSI design","score":0.004100000020116568,"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/T11697","display_name":"Numerical Methods and Algorithms","score":0.0031999999191612005,"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"}}],"keywords":[{"id":"https://openalex.org/keywords/activation-quantization","display_name":"activation quantization","score":0.8069999814033508},{"id":"https://openalex.org/keywords/multimodal-large-language-models","display_name":"multimodal large language models","score":0.9300000071525574},{"id":"https://openalex.org/keywords/low-bit-quantization","display_name":"low-bit quantization","score":0.9190000295639038},{"id":"https://openalex.org/keywords/quantization-error-compensation","display_name":"quantization error compensation","score":0.878000020980835},{"id":"https://openalex.org/keywords/video-generation","display_name":"video generation","score":0.8090000152587891},{"id":"https://openalex.org/keywords/multimodal-reasoning","display_name":"multimodal reasoning","score":0.7319999933242798},{"id":"https://openalex.org/keywords/qwen3-vl","display_name":"Qwen3-VL","score":0.7289999723434448}],"concepts":[{"id":"https://openalex.org/C28855332","wikidata":"https://www.wikidata.org/wiki/Q198099","display_name":"Quantization (signal processing)","level":2,"score":0.8363000154495239},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.6523000001907349},{"id":"https://openalex.org/C155512373","wikidata":"https://www.wikidata.org/wiki/Q287450","display_name":"Residual","level":2,"score":0.5882999897003174},{"id":"https://openalex.org/C79337645","wikidata":"https://www.wikidata.org/wiki/Q779824","display_name":"Outlier","level":2,"score":0.5651000142097473},{"id":"https://openalex.org/C199833920","wikidata":"https://www.wikidata.org/wiki/Q612536","display_name":"Vector quantization","level":2,"score":0.4560000002384186},{"id":"https://openalex.org/C136625980","wikidata":"https://www.wikidata.org/wiki/Q663208","display_name":"Rounding","level":2,"score":0.41760000586509705},{"id":"https://openalex.org/C11413529","wikidata":"https://www.wikidata.org/wiki/Q8366","display_name":"Algorithm","level":1,"score":0.4047999978065491},{"id":"https://openalex.org/C2776459999","wikidata":"https://www.wikidata.org/wiki/Q2119376","display_name":"Fidelity","level":2,"score":0.400299996137619},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.3846000134944916},{"id":"https://openalex.org/C2780513914","wikidata":"https://www.wikidata.org/wiki/Q18210350","display_name":"Bottleneck","level":2,"score":0.3716999888420105},{"id":"https://openalex.org/C115051666","wikidata":"https://www.wikidata.org/wiki/Q6522493","display_name":"Ranging","level":2,"score":0.3163999915122986},{"id":"https://openalex.org/C113775141","wikidata":"https://www.wikidata.org/wiki/Q428691","display_name":"Computer engineering","level":1,"score":0.29580000042915344},{"id":"https://openalex.org/C68339613","wikidata":"https://www.wikidata.org/wiki/Q1549489","display_name":"Speedup","level":2,"score":0.295199990272522},{"id":"https://openalex.org/C179799912","wikidata":"https://www.wikidata.org/wiki/Q205084","display_name":"Computational complexity theory","level":2,"score":0.295199990272522},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.2809999883174896},{"id":"https://openalex.org/C45374587","wikidata":"https://www.wikidata.org/wiki/Q12525525","display_name":"Computation","level":2,"score":0.27889999747276306}],"mesh":[],"locations_count":2,"locations":[{"id":"pmh:oai:arXiv.org:2608.26581","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2608.26581","pdf_url":"https://arxiv.org/pdf/2608.26581","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":"public-domain","license_id":"https://openalex.org/licenses/public-domain","version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},{"id":"doi:10.48550/arxiv.2608.26581","doi":"https://doi.org/10.48550/arxiv.2608.26581","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2608.26581","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":"public-domain","license_id":"https://openalex.org/licenses/public-domain","version":null,"is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"Preprint"}],"best_oa_location":{"id":"pmh:oai:arXiv.org:2608.26581","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2608.26581","pdf_url":"https://arxiv.org/pdf/2608.26581","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":"public-domain","license_id":"https://openalex.org/licenses/public-domain","version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},"sustainable_development_goals":[],"study_designs":[],"awards":[],"funders":[],"has_content":{"grobid_xml":true,"pdf":true},"content_urls":{"pdf":"https://content.openalex.org/works/W7204589553.pdf","grobid_xml":"https://content.openalex.org/works/W7204589553.grobid-xml"},"referenced_works_count":18,"referenced_works":["https://openalex.org/W2809624076","https://openalex.org/W4307934016","https://openalex.org/W4387725429","https://openalex.org/W4400375159","https://openalex.org/W4403580037","https://openalex.org/W4407310230","https://openalex.org/W4415332678","https://openalex.org/W4415798312","https://openalex.org/W4416544157","https://openalex.org/W4416780398","https://openalex.org/W4417113160","https://openalex.org/W4417536752","https://openalex.org/W7077077575","https://openalex.org/W7128775491","https://openalex.org/W7133198810","https://openalex.org/W7133333059","https://openalex.org/W7148297826","https://openalex.org/W7154081199"],"related_works":["https://openalex.org/W4415332678","https://openalex.org/W4402897324","https://openalex.org/W7163911910","https://openalex.org/W4416441354","https://openalex.org/W7214740394","https://openalex.org/W4389519530","https://openalex.org/W4387635190","https://openalex.org/W7162556827","https://openalex.org/W6966407491","https://openalex.org/W6947648164"],"abstract_inverted_index":{"Low-bit":[0],"quantization":[1,55,79,90,139,183,206],"offers":[2],"a":[3,50,115,174],"promising":[4],"avenue":[5],"for":[6,22,138,225],"reducing":[7],"the":[8,92,122,147,178,182,190,208],"computational":[9,161],"and":[10,34,43,65,136,158,167,186,198,214,222],"memory":[11],"demands":[12],"of":[13,53,95,150,177,184],"Multimodal":[14],"Large":[15],"Language":[16],"Models":[17],"(MLLMs).":[18],"Recent":[19],"hardware":[20],"support":[21],"low-precision":[23],"formats,":[24],"ranging":[25],"from":[26],"MXFP8":[27,72],"to":[28,81,192],"ultra-low-bit":[29,151,212],"formats":[30],"such":[31],"as":[32,91,207,219],"MXFP4":[33,185],"HiF4,":[35,187],"has":[36],"accelerated":[37],"research":[38],"into":[39],"efficient":[40],"MLLM":[41],"training":[42],"deployment.":[44,228],"In":[45],"this":[46,96,107],"work,":[47],"we":[48,87,109],"present":[49],"systematic":[51],"study":[52],"these":[54],"schemes":[56],"in":[57,211],"representative":[58],"MLLMs":[59,213],"that":[60,71,120,170],"span":[61],"both":[62,196],"video":[63],"generation":[64,197],"reasoning":[66,200],"tasks.":[67],"Our":[68,202],"analysis":[69],"shows":[70],"achieves":[73],"near-lossless":[74],"performance,":[75],"whereas":[76],"aggressive":[77],"4-bit":[78,227],"leads":[80],"significant":[82],"degradation.":[83],"Through":[84],"extensive":[85],"ablations,":[86],"identify":[88],"activation":[89,117,125,143,205,217],"primary":[93,123],"source":[94],"performance":[97,179],"loss,":[98],"contributing":[99],"substantially":[100],"more":[101],"than":[102],"weight":[103],"quantization.":[104],"Motivated":[105],"by":[106],"observation,":[108],"propose":[110],"Residual":[111],"Fallback":[112],"Quantization":[113],"(RFQ),":[114],"lightweight":[116],"reconstruction":[118,218],"framework":[119],"supplements":[121],"ulta-low-bit":[124],"representation":[126],"with":[127],"an":[128,220],"auxiliary":[129],"quantized":[130],"residual":[131],"pathway.":[132],"By":[133],"explicitly":[134],"modeling":[135],"compensating":[137],"errors,":[140],"RFQ":[141,153,171],"improves":[142],"fidelity":[144],"while":[145],"preserving":[146],"efficiency":[148],"advantages":[149],"computation.":[152],"requires":[154],"no":[155],"architectural":[156],"modifications":[157],"incurs":[159],"negligible":[160],"overhead.":[162],"Extensive":[163],"experiments":[164],"on":[165],"Wan2.2":[166],"Qwen3-VL":[168],"demonstrate":[169],"consistently":[172],"recovers":[173],"substantial":[175],"portion":[176],"lost":[180],"under":[181],"significantly":[188],"narrowing":[189],"gap":[191],"BF16":[193],"baselines":[194],"across":[195],"4":[199],"benchmarks.":[201],"findings":[203],"establish":[204],"dominant":[209],"bottleneck":[210],"highlight":[215],"residual-based":[216],"effective":[221],"practical":[223],"strategy":[224],"robust":[226]},"counts_by_year":[],"updated_date":"2026-11-07T11:25:07.637180","created_date":"2026-08-29T00:00:00"}