{"id":"https://openalex.org/W7164591065","doi":"https://doi.org/10.48550/arxiv.2606.12876","title":"Multi-Bitwidth Quantization for LLMs Using Additive Codebooks","display_name":"Multi-Bitwidth Quantization for LLMs Using Additive Codebooks","publication_year":2026,"publication_date":"2026-06-11","ids":{"openalex":"https://openalex.org/W7164591065","doi":"https://doi.org/10.48550/arxiv.2606.12876"},"language":"en","primary_location":{"id":"pmh:oai:arXiv.org:2606.12876","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2606.12876","pdf_url":"https://arxiv.org/pdf/2606.12876","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":"cc-by","license_id":"https://openalex.org/licenses/cc-by","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/2606.12876","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5036000605","display_name":"Liza Babaoglu","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I185261750","display_name":"University of Toronto","ror":"https://ror.org/03dbr7087","country_code":"CA","type":"education","lineage":["https://openalex.org/I185261750"]}],"countries":["CA"],"is_corresponding":false,"raw_author_name":"Babaoglu, Liza","raw_affiliation_strings":["Department of Electrical and Computer Engineering , University of Toronto"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Department of Electrical and Computer Engineering , University of Toronto","institution_ids":["https://openalex.org/I185261750"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5101203523","display_name":"Shuangyi Chen","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I185261750","display_name":"University of Toronto","ror":"https://ror.org/03dbr7087","country_code":"CA","type":"education","lineage":["https://openalex.org/I185261750"]}],"countries":["CA"],"is_corresponding":false,"raw_author_name":"Chen, Shuangyi","raw_affiliation_strings":["Department of Electrical and Computer Engineering , University of Toronto"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Department of Electrical and Computer Engineering , University of Toronto","institution_ids":["https://openalex.org/I185261750"]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5046253142","display_name":"Ashish J. Khisti","orcid":"https://orcid.org/0000-0002-2331-8965","observed_orcids":["https://orcid.org/0000-0002-2331-8965"]},"institutions":[{"id":"https://openalex.org/I185261750","display_name":"University of Toronto","ror":"https://ror.org/03dbr7087","country_code":"CA","type":"education","lineage":["https://openalex.org/I185261750"]}],"countries":["CA"],"is_corresponding":false,"raw_author_name":"Khisti, Ashish","raw_affiliation_strings":["Department of Electrical and Computer Engineering , University of Toronto"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Department of Electrical and Computer Engineering , University of Toronto","institution_ids":["https://openalex.org/I185261750"]}]}],"institutions":[],"countries_distinct_count":1,"institutions_distinct_count":1,"corresponding_author_ids":[],"corresponding_institution_ids":["https://openalex.org/I185261750"],"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":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":0.9409000277519226,"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":0.9409000277519226,"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/T10028","display_name":"Topic Modeling","score":0.009700000286102295,"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/T10901","display_name":"Advanced Data Compression Techniques","score":0.00930000003427267,"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"}}],"keywords":[{"id":"https://openalex.org/keywords/large-language-models","display_name":"large language models","score":0.949999988079071},{"id":"https://openalex.org/keywords/post-training-quantization","display_name":"post-training quantization","score":0.8550000190734863},{"id":"https://openalex.org/keywords/successive-refinement","display_name":"successive refinement","score":0.8040000200271606},{"id":"https://openalex.org/keywords/weighted-mean-squared-error","display_name":"weighted mean squared error","score":0.7860000133514404},{"id":"https://openalex.org/keywords/decoding-latency-reduction","display_name":"decoding latency reduction","score":0.6380000114440918}],"concepts":[{"id":"https://openalex.org/C28855332","wikidata":"https://www.wikidata.org/wiki/Q198099","display_name":"Quantization (signal processing)","level":2,"score":0.6459000110626221},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.6004999876022339},{"id":"https://openalex.org/C11413529","wikidata":"https://www.wikidata.org/wiki/Q8366","display_name":"Algorithm","level":1,"score":0.5612000226974487},{"id":"https://openalex.org/C100279451","wikidata":"https://www.wikidata.org/wiki/Q372193","display_name":"Perplexity","level":3,"score":0.5034999847412109},{"id":"https://openalex.org/C127759330","wikidata":"https://www.wikidata.org/wiki/Q637416","display_name":"Codebook","level":2,"score":0.44859999418258667},{"id":"https://openalex.org/C139945424","wikidata":"https://www.wikidata.org/wiki/Q1940696","display_name":"Mean squared error","level":2,"score":0.3878999948501587},{"id":"https://openalex.org/C2776459999","wikidata":"https://www.wikidata.org/wiki/Q2119376","display_name":"Fidelity","level":2,"score":0.387800008058548},{"id":"https://openalex.org/C2779960059","wikidata":"https://www.wikidata.org/wiki/Q7113681","display_name":"Overhead (engineering)","level":2,"score":0.3862000107765198},{"id":"https://openalex.org/C137293760","wikidata":"https://www.wikidata.org/wiki/Q3621696","display_name":"Language model","level":2,"score":0.3804999887943268},{"id":"https://openalex.org/C163716315","wikidata":"https://www.wikidata.org/wiki/Q901177","display_name":"Gaussian","level":2,"score":0.35089999437332153},{"id":"https://openalex.org/C61326573","wikidata":"https://www.wikidata.org/wiki/Q1496376","display_name":"Gaussian process","level":3,"score":0.33719998598098755},{"id":"https://openalex.org/C77618280","wikidata":"https://www.wikidata.org/wiki/Q1155772","display_name":"Scheme (mathematics)","level":2,"score":0.3255999982357025},{"id":"https://openalex.org/C33923547","wikidata":"https://www.wikidata.org/wiki/Q395","display_name":"Mathematics","level":0,"score":0.3093000054359436},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.29120001196861267},{"id":"https://openalex.org/C64185310","wikidata":"https://www.wikidata.org/wiki/Q843483","display_name":"Rate\u2013distortion theory","level":3,"score":0.28949999809265137},{"id":"https://openalex.org/C80444323","wikidata":"https://www.wikidata.org/wiki/Q2878974","display_name":"Theoretical computer science","level":1,"score":0.2863999903202057},{"id":"https://openalex.org/C126780896","wikidata":"https://www.wikidata.org/wiki/Q899871","display_name":"Distortion (music)","level":4,"score":0.272599995136261},{"id":"https://openalex.org/C2780898871","wikidata":"https://www.wikidata.org/wiki/Q860554","display_name":"Performance metric","level":2,"score":0.259799987077713},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.25929999351501465},{"id":"https://openalex.org/C199833920","wikidata":"https://www.wikidata.org/wiki/Q612536","display_name":"Vector quantization","level":2,"score":0.25870001316070557},{"id":"https://openalex.org/C155512373","wikidata":"https://www.wikidata.org/wiki/Q287450","display_name":"Residual","level":2,"score":0.25429999828338623}],"mesh":[],"locations_count":2,"locations":[{"id":"pmh:oai:arXiv.org:2606.12876","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2606.12876","pdf_url":"https://arxiv.org/pdf/2606.12876","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":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},{"id":"doi:10.48550/arxiv.2606.12876","doi":"https://doi.org/10.48550/arxiv.2606.12876","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2606.12876","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":"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"}],"best_oa_location":{"id":"pmh:oai:arXiv.org:2606.12876","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2606.12876","pdf_url":"https://arxiv.org/pdf/2606.12876","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":"cc-by","license_id":"https://openalex.org/licenses/cc-by","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/W7164591065.pdf","grobid_xml":"https://content.openalex.org/works/W7164591065.grobid-xml"},"referenced_works_count":15,"referenced_works":["https://openalex.org/W2525332836","https://openalex.org/W2794325560","https://openalex.org/W4380714727","https://openalex.org/W4390833798","https://openalex.org/W4391940867","https://openalex.org/W4392182282","https://openalex.org/W4392822465","https://openalex.org/W4401306886","https://openalex.org/W4405655184","https://openalex.org/W4407359588","https://openalex.org/W4415249820","https://openalex.org/W7127604893","https://openalex.org/W7135012008","https://openalex.org/W7164150941","https://openalex.org/W7164197693"],"related_works":["https://openalex.org/W7163591689","https://openalex.org/W7161147056","https://openalex.org/W7196978668","https://openalex.org/W4409362928","https://openalex.org/W7127571195","https://openalex.org/W4378770729","https://openalex.org/W4387635190","https://openalex.org/W4416035853","https://openalex.org/W7131393438","https://openalex.org/W4399061648"],"abstract_inverted_index":{"As":[0,48],"large":[1],"language":[2],"models":[3],"(LLMs)":[4],"are":[5,56],"increasingly":[6],"deployed":[7],"across":[8,73],"heterogeneous":[9],"hardware":[10],"with":[11,169],"varying":[12],"resource":[13],"constraints,":[14],"the":[15,20,122,151],"ability":[16],"to":[17,81,126,150],"adaptively":[18],"manage":[19],"performance-efficiency":[21],"trade-off":[22],"without":[23],"retraining":[24],"is":[25,133],"critical.":[26],"We":[27],"propose":[28],"Drop-by-Drop,":[29],"a":[30,44,60,118,134],"novel":[31],"multi-bitwidth":[32,167],"post-training":[33],"quantization":[34,147],"framework":[35],"enabling":[36],"inference-time":[37],"precision":[38,115],"control":[39],"over":[40],"LLM":[41,68],"weights":[42],"from":[43],"single":[45,135],"quantized":[46],"model.":[47],"theoretical":[49],"motivation,":[50],"we":[51],"prove":[52],"that":[53,137],"Gaussian":[54,128],"sources":[55],"successively":[57],"refinable":[58],"under":[59],"weighted":[61],"mean":[62],"squared":[63],"error":[64],"distortion":[65,74],"motivated":[66],"by":[67,94,141],"loss":[69],"functions:":[70],"successive":[71],"refinement":[72],"levels":[75],"incurs":[76],"no":[77],"rate":[78],"penalty":[79],"relative":[80,149],"encoding":[82],"at":[83,113],"each":[84,114],"level":[85],"separately.":[86],"Drop-by-Drop":[87,161],"approximates":[88],"this":[89],"hierarchical":[90],"structure":[91],"in":[92,105],"practice":[93],"incorporating":[95],"Matryoshka-style":[96],"supervision":[97],"into":[98],"additive":[99],"codebook":[100,107],"training,":[101],"inducing":[102],"an":[103],"ordering":[104],"which":[106],"prefixes":[108],"yield":[109],"accurate":[110],"partial":[111],"reconstructions":[112],"level.":[116],"Furthermore,":[117],"block-Hadamard":[119],"rotation":[120],"brings":[121],"weight":[123],"distribution":[124],"closer":[125],"our":[127,174],"source":[129],"assumption.":[130],"The":[131],"result":[132],"model":[136],"serves":[138],"multiple":[139],"bitwidths":[140],"dropping":[142],"codebooks,":[143],"reducing":[144],"storage":[145],"and":[146,159],"cost":[148],"static":[152],"per-bitwidth":[153],"models.":[154],"Across":[155],"Qwen,":[156],"LLaMA,":[157],"Gemma,":[158],"Mistral,":[160],"achieves":[162],"lower":[163],"perplexity":[164],"than":[165],"state-of-the-art":[166],"methods":[168],"competitive":[170],"zero-shot":[171],"accuracy,":[172],"while":[173],"specialized":[175],"kernels":[176],"further":[177],"reduce":[178],"decoding":[179],"latency.":[180]},"counts_by_year":[{"year":2026,"cited_by_count":1}],"updated_date":"2026-11-08T12:12:28.789561","created_date":"2026-06-13T00:00:00"}