{"id":"https://openalex.org/W7214733366","doi":"https://doi.org/10.48550/arxiv.2609.34866","title":"From Attention Sensitivity to Layer Role: Revisiting Mixed-Precision Quantization of Transformers","display_name":"From Attention Sensitivity to Layer Role: Revisiting Mixed-Precision Quantization of Transformers","publication_year":2026,"publication_date":"2026-09-28","ids":{"openalex":"https://openalex.org/W7214733366","doi":"https://doi.org/10.48550/arxiv.2609.34866"},"language":"en","primary_location":{"id":"pmh:oai:arXiv.org:2609.34866","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2609.34866","pdf_url":"https://arxiv.org/pdf/2609.34866","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/2609.34866","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5116293366","display_name":"Nafiseh HosseinpourFardi","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I133529467","display_name":"Sharif University of Technology","ror":"https://ror.org/024c2fq17","country_code":"IR","type":"education","lineage":["https://openalex.org/I133529467"]}],"countries":["IR"],"is_corresponding":false,"raw_author_name":"HosseinpourFardi, Nafiseh","raw_affiliation_strings":["Sharif University of Technology , Tehran , Iran"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Sharif University of Technology , Tehran , Iran","institution_ids":["https://openalex.org/I133529467"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5153450888","display_name":"Negar Alihadi","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I133529467","display_name":"Sharif University of Technology","ror":"https://ror.org/024c2fq17","country_code":"IR","type":"education","lineage":["https://openalex.org/I133529467"]}],"countries":["IR"],"is_corresponding":false,"raw_author_name":"Alihadi, Negar","raw_affiliation_strings":["Sharif University of Technology , Tehran , Iran"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Sharif University of Technology , Tehran , Iran","institution_ids":["https://openalex.org/I133529467"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5033317835","display_name":"Mahmoudreza Babaei","orcid":"https://orcid.org/0000-0003-1141-9758","observed_orcids":["https://orcid.org/0000-0003-1141-9758"]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Babaei, Mahmoudreza","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5128029756","display_name":"Milad Hosseini","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I4387153096","display_name":"Computer Vision Center","ror":"https://ror.org/00s0nnj93","country_code":"ES","type":"other","lineage":["https://openalex.org/I4387153040","https://openalex.org/I4387153096"]}],"countries":["ES"],"is_corresponding":false,"raw_author_name":"Hosseini, Milad","raw_affiliation_strings":["Computer Vision Center (CVC) , Barcelona , Spain"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Computer Vision Center (CVC) , Barcelona , Spain","institution_ids":["https://openalex.org/I4387153096"]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5042278493","display_name":"Adrian Weller","orcid":"https://orcid.org/0000-0003-1915-7158","observed_orcids":["https://orcid.org/0000-0003-1915-7158"]},"institutions":[{"id":"https://openalex.org/I241749","display_name":"University of Cambridge","ror":"https://ror.org/013meh722","country_code":"GB","type":"education","lineage":["https://openalex.org/I241749"]}],"countries":["GB"],"is_corresponding":false,"raw_author_name":"Weller, Adrian","raw_affiliation_strings":["University of Cambridge , United Kingdom"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"University of Cambridge , United Kingdom","institution_ids":["https://openalex.org/I241749"]}]}],"institutions":[],"countries_distinct_count":3,"institutions_distinct_count":3,"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":0.9667999744415283,"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":0.9667999744415283,"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":0.003700000001117587,"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/T11307","display_name":"Domain Adaptation and Few-Shot Learning","score":0.00279999990016222,"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/mixed-precision-quantization","display_name":"mixed-precision quantization","score":0.9279999732971191},{"id":"https://openalex.org/keywords/transformer","display_name":"transformer models","score":0.6669999957084656},{"id":"https://openalex.org/keywords/attention-layers","display_name":"attention layers","score":0.5059999823570251},{"id":"https://openalex.org/keywords/post-training-quantization","display_name":"post-training quantization","score":0.8479999899864197},{"id":"https://openalex.org/keywords/multi-bit-quantization","display_name":"multi-bit quantization","score":0.5440000295639038},{"id":"https://openalex.org/keywords/mistral-7b","display_name":"Mistral-7B","score":0.7699999809265137},{"id":"https://openalex.org/keywords/gpt-2","display_name":"GPT-2","score":0.8209999799728394},{"id":"https://openalex.org/keywords/fine-tuning","display_name":"fine-tuning","score":0.492000013589859},{"id":"https://openalex.org/keywords/model-compression","display_name":"model compression","score":0.46700000762939453}],"concepts":[{"id":"https://openalex.org/C28855332","wikidata":"https://www.wikidata.org/wiki/Q198099","display_name":"Quantization (signal processing)","level":2,"score":0.6696000099182129},{"id":"https://openalex.org/C33923547","wikidata":"https://www.wikidata.org/wiki/Q395","display_name":"Mathematics","level":0,"score":0.5321999788284302},{"id":"https://openalex.org/C11413529","wikidata":"https://www.wikidata.org/wiki/Q8366","display_name":"Algorithm","level":1,"score":0.5259000062942505},{"id":"https://openalex.org/C175291020","wikidata":"https://www.wikidata.org/wiki/Q1156822","display_name":"Offset (computer science)","level":2,"score":0.4381999969482422},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.391400009393692},{"id":"https://openalex.org/C45374587","wikidata":"https://www.wikidata.org/wiki/Q12525525","display_name":"Computation","level":2,"score":0.38909998536109924},{"id":"https://openalex.org/C21200559","wikidata":"https://www.wikidata.org/wiki/Q7451068","display_name":"Sensitivity (control systems)","level":2,"score":0.3693999946117401},{"id":"https://openalex.org/C206588197","wikidata":"https://www.wikidata.org/wiki/Q846574","display_name":"Reuse","level":2,"score":0.361299991607666},{"id":"https://openalex.org/C63479239","wikidata":"https://www.wikidata.org/wiki/Q7353546","display_name":"Robustness (evolution)","level":3,"score":0.30630001425743103},{"id":"https://openalex.org/C55166926","wikidata":"https://www.wikidata.org/wiki/Q2892946","display_name":"Oracle","level":2,"score":0.3000999987125397},{"id":"https://openalex.org/C177774035","wikidata":"https://www.wikidata.org/wiki/Q1246948","display_name":"Granularity","level":2,"score":0.2989000082015991},{"id":"https://openalex.org/C106487976","wikidata":"https://www.wikidata.org/wiki/Q685816","display_name":"Matrix (chemical analysis)","level":2,"score":0.2944999933242798},{"id":"https://openalex.org/C126255220","wikidata":"https://www.wikidata.org/wiki/Q141495","display_name":"Mathematical optimization","level":1,"score":0.26919999718666077},{"id":"https://openalex.org/C2779417484","wikidata":"https://www.wikidata.org/wiki/Q406009","display_name":"Dart","level":2,"score":0.2572999894618988},{"id":"https://openalex.org/C2777210771","wikidata":"https://www.wikidata.org/wiki/Q4927124","display_name":"Block (permutation group theory)","level":2,"score":0.2556000053882599},{"id":"https://openalex.org/C2780069185","wikidata":"https://www.wikidata.org/wiki/Q7977945","display_name":"Equivalence (formal languages)","level":2,"score":0.251800000667572},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.2506999969482422},{"id":"https://openalex.org/C48145219","wikidata":"https://www.wikidata.org/wiki/Q1335365","display_name":"Security token","level":2,"score":0.25060001015663147}],"mesh":[],"locations_count":2,"locations":[{"id":"pmh:oai:arXiv.org:2609.34866","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2609.34866","pdf_url":"https://arxiv.org/pdf/2609.34866","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.2609.34866","doi":"https://doi.org/10.48550/arxiv.2609.34866","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2609.34866","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:2609.34866","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2609.34866","pdf_url":"https://arxiv.org/pdf/2609.34866","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":{"pdf":true,"grobid_xml":true},"content_urls":{"pdf":"https://content.openalex.org/works/W7214733366.pdf","grobid_xml":"https://content.openalex.org/works/W7214733366.grobid-xml"},"referenced_works_count":7,"referenced_works":["https://openalex.org/W2242818861","https://openalex.org/W4386231419","https://openalex.org/W4387561528","https://openalex.org/W4393147284","https://openalex.org/W4413025361","https://openalex.org/W4417349452","https://openalex.org/W7079896301"],"related_works":["https://openalex.org/W7214108747","https://openalex.org/W7171703249","https://openalex.org/W7214228266","https://openalex.org/W7105211489","https://openalex.org/W7125218531","https://openalex.org/W7167073937","https://openalex.org/W7213402146","https://openalex.org/W7163026552","https://openalex.org/W7171523575","https://openalex.org/W7220923247"],"abstract_inverted_index":{"Most":[0],"post-training":[1],"quantization":[2,117],"pipelines":[3],"fit":[4,93],"each":[5],"weight":[6,223],"matrix":[7,13,221],"to":[8,92,106,148,195,291],"its":[9,138],"pretrained":[10,279],"counterpart,":[11],"one":[12,67,248],"at":[14,57,172,200,209,216],"a":[15,78,111,150,153,186,222,251],"time.":[16],"Whether":[17],"that":[18],"proxy":[19,244],"tracks":[20,141],"what":[21],"an":[22,142,242],"attention":[23,50,86,283,293],"block":[24,109],"actually":[25],"computes,":[26],"or":[27],"how":[28],"errors":[29],"in":[30,225],"the":[31,39,46,49,63,71,82,94,108,129,162,181,193,217],"Q,":[32,73],"K":[33],"and":[34,59,88,105,134,137,179],"V":[35,75],"projections":[36,56],"compound":[37],"inside":[38],"softmax,":[40],"is":[41,241,261,266],"rarely":[42],"checked.":[43],"We":[44],"write":[45],"objective":[47,254],"on":[48,176,180,270],"output":[51],"instead,":[52],"over":[53,70],"all":[54,173],"three":[55],"once,":[58],"reuse":[60],"it":[61,90,189],"throughout":[62],"pipeline.":[64],"JAB":[65,125,175,215],"defines":[66],"scalar":[68],"loss":[69],"joint":[72],"K,":[74],"weights":[76,96,194,275],"of":[77,118,128,152,192,205],"block,":[79],"evaluated":[80],"against":[81,212],"block's":[83,252],"real":[84],"causally-masked":[85],"output,":[87],"uses":[89],"twice:":[91],"quantized":[95],"(GPTQ":[97],"warm":[98],"start,":[99],"then":[100],"STE":[101],"with":[102,185,286],"learnable":[103],"scales),":[104],"score":[107],"for":[110,214,245],"multiple-choice":[112],"knapsack":[113],"allocation.":[114,163],"On":[115],"attention-only":[116,271],"Mistral-7B":[119,183],"this":[120],"works.":[121],"At":[122],"3":[123],"bits":[124,197],"recovers":[126],"77-90%":[127],"gap":[130],"between":[131],"uniform":[132],"GPTQ":[133],"full":[135,182,206],"precision,":[136],"sensitivity":[139,170,230],"estimate":[140,171,231],"oracle":[143],"costing":[144],"73":[145],"forward":[146],"passes":[147],"within":[149,203],"fraction":[151],"point.":[154],"It":[155],"stops":[156],"working":[157],"once":[158],"MLP":[159,178],"layers":[160],"enter":[161],"A":[164],"role-aware":[165],"offset":[166],"rule":[167],"needing":[168],"no":[169],"beats":[174],"GPT-2's":[177],"model:":[184],"3-bit":[187],"floor":[188],"quantizes":[190],"96.4%":[191],"4.5":[196],"per":[198],"parameter":[199],"6.933":[201],"perplexity,":[202],"4.4%":[204],"precision":[207],"(6.643)":[208],"3.56x":[210],"compression,":[211],"7.158":[213],"same":[218],"budget.":[219],"Which":[220],"sits":[224],"matters":[226],"more":[227],"than":[228],"any":[229],"we":[232],"computed.":[233],"Two":[234],"things":[235],"came":[236],"out":[237],"sideways.":[238],"Block-local":[239],"reconstruction":[240],"unreliable":[243],"end-to-end":[246],"perplexity:":[247],"run":[249],"improved":[250],"own":[253],"4.6x":[255],"while":[256,281],"perplexity":[257],"rose":[258],"32x,":[259],"which":[260],"why":[262],"every":[263],"allocation":[264],"here":[265],"validated":[267],"end-to-end.":[268],"And":[269],"quantization,":[272],"fine-tuning":[273],"moved":[274],"farther":[276],"from":[277],"their":[278],"values":[280],"pulling":[282],"outputs":[284],"closer,":[285],"net":[287],"gains.":[288],"Post-training":[289],"seems":[290],"recover":[292],"behavior,":[294],"not":[295],"weights.":[296]},"counts_by_year":[],"updated_date":"2026-11-10T11:39:43.021295","created_date":"2026-09-30T00:00:00"}