{"id":"https://openalex.org/W7167073937","doi":"https://doi.org/10.48550/arxiv.2607.00908","title":"Beyond Activation Alignment:The Alignment-Diversity Tradeoff in Task-Aware LLM Quantization","display_name":"Beyond Activation Alignment:The Alignment-Diversity Tradeoff in Task-Aware LLM Quantization","publication_year":2026,"publication_date":"2026-07-01","ids":{"openalex":"https://openalex.org/W7167073937","doi":"https://doi.org/10.48550/arxiv.2607.00908"},"language":null,"primary_location":{"id":"doi:10.48550/arxiv.2607.00908","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2607.00908","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,"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.2607.00908","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5029787563","display_name":"Fang Wang","orcid":"https://orcid.org/0000-0002-3327-4177"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Wang, Fei","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5139853425","display_name":"Chao Xue","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Xue, Chao","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5087289491","display_name":"Taoran Liu","orcid":"https://orcid.org/0000-0002-7806-1263"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Liu, Taoran","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5139943399","display_name":"Li Shen","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Shen, Li","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5139915775","display_name":"Ye Liu","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Liu, Ye","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5038748720","display_name":"Changxing Ding","orcid":"https://orcid.org/0000-0001-7232-3181"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Ding, ChangXing","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/T10181","display_name":"Natural Language Processing Techniques","score":0.3328000009059906,"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"}},"topics":[{"id":"https://openalex.org/T10181","display_name":"Natural Language Processing Techniques","score":0.3328000009059906,"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":0.1412000060081482,"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/T10028","display_name":"Topic Modeling","score":0.09149999916553497,"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/perplexity","display_name":"Perplexity","score":0.7114999890327454},{"id":"https://openalex.org/keywords/robustness","display_name":"Robustness (evolution)","score":0.6391000151634216},{"id":"https://openalex.org/keywords/sensitivity","display_name":"Sensitivity (control systems)","score":0.5630999803543091},{"id":"https://openalex.org/keywords/quantization","display_name":"Quantization (signal processing)","score":0.5181000232696533},{"id":"https://openalex.org/keywords/correlation","display_name":"Correlation","score":0.43220001459121704},{"id":"https://openalex.org/keywords/language-model","display_name":"Language model","score":0.4203999936580658}],"concepts":[{"id":"https://openalex.org/C100279451","wikidata":"https://www.wikidata.org/wiki/Q372193","display_name":"Perplexity","level":3,"score":0.7114999890327454},{"id":"https://openalex.org/C63479239","wikidata":"https://www.wikidata.org/wiki/Q7353546","display_name":"Robustness (evolution)","level":3,"score":0.6391000151634216},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.5637000203132629},{"id":"https://openalex.org/C21200559","wikidata":"https://www.wikidata.org/wiki/Q7451068","display_name":"Sensitivity (control systems)","level":2,"score":0.5630999803543091},{"id":"https://openalex.org/C28855332","wikidata":"https://www.wikidata.org/wiki/Q198099","display_name":"Quantization (signal processing)","level":2,"score":0.5181000232696533},{"id":"https://openalex.org/C117220453","wikidata":"https://www.wikidata.org/wiki/Q5172842","display_name":"Correlation","level":2,"score":0.43220001459121704},{"id":"https://openalex.org/C11413529","wikidata":"https://www.wikidata.org/wiki/Q8366","display_name":"Algorithm","level":1,"score":0.4260999858379364},{"id":"https://openalex.org/C137293760","wikidata":"https://www.wikidata.org/wiki/Q3621696","display_name":"Language model","level":2,"score":0.4203999936580658},{"id":"https://openalex.org/C4679612","wikidata":"https://www.wikidata.org/wiki/Q866298","display_name":"Aggregate (composite)","level":2,"score":0.3905999958515167},{"id":"https://openalex.org/C33923547","wikidata":"https://www.wikidata.org/wiki/Q395","display_name":"Mathematics","level":0,"score":0.3808000087738037},{"id":"https://openalex.org/C41426520","wikidata":"https://www.wikidata.org/wiki/Q1192065","display_name":"Point estimation","level":2,"score":0.3797000050544739},{"id":"https://openalex.org/C165838908","wikidata":"https://www.wikidata.org/wiki/Q736777","display_name":"Calibration","level":2,"score":0.3398999869823456},{"id":"https://openalex.org/C129848803","wikidata":"https://www.wikidata.org/wiki/Q2564360","display_name":"Sample size determination","level":2,"score":0.3003000020980835},{"id":"https://openalex.org/C83546350","wikidata":"https://www.wikidata.org/wiki/Q1139051","display_name":"Regression","level":2,"score":0.2973000109195709},{"id":"https://openalex.org/C61797465","wikidata":"https://www.wikidata.org/wiki/Q1188986","display_name":"Term (time)","level":2,"score":0.2840999960899353},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.2824000120162964},{"id":"https://openalex.org/C105795698","wikidata":"https://www.wikidata.org/wiki/Q12483","display_name":"Statistics","level":1,"score":0.2741999924182892}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.48550/arxiv.2607.00908","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2607.00908","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,"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.2607.00908","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2607.00908","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,"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":{"grobid_xml":false,"pdf":false},"content_urls":null,"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Mixed-precision":[0],"quantization":[1],"(MPQ)":[2],"has":[3],"become":[4],"a":[5,22,95,111,115,142,203],"key":[6],"technique":[7],"for":[8,47,110],"deploying":[9],"large":[10],"language":[11],"models":[12,148],"under":[13],"stringent":[14],"memory":[15],"and":[16,80,103,120,124,132,139,176],"compute":[17],"constraints.":[18],"We":[19,58],"first":[20],"identify":[21],"phenomenon":[23],"that":[24,43,98,196],"we":[25,89],"term":[26],"the":[27,179],"Perplexity":[28],"Illusion:":[29],"layers":[30],"ranked":[31],"as":[32],"important":[33],"by":[34,185],"perplexity-based":[35],"sensitivity":[36,78,126],"show":[37,195],"little":[38],"rank":[39],"correlation":[40],"with":[41,51],"those":[42],"are":[44],"most":[45],"influential":[46],"complex":[48],"reasoning":[49],"performance,":[50,72],"Kendall":[52],"$\u03c4\\approx":[53],"0$":[54],"in":[55,173,206],"our":[56],"analysis.":[57],"further":[59],"reveal":[60,141],"an":[61,158],"Alignment-Diversity":[62],"Tradeoff:":[63],"using":[64,114],"only":[65],"target-task":[66],"calibration":[67],"data":[68,76],"can":[69,149],"degrade":[70],"post-quantization":[71],"whereas":[73],"incorporating":[74],"general-domain":[75],"stabilizes":[77],"estimation":[79],"improves":[81,177],"robustness":[82],"across":[83],"tasks.":[84],"Based":[85],"on":[86,137,183,191],"these":[87],"observations,":[88],"propose":[90],"TASA":[91,108,164],"(Task-Aware":[92],"Sensitivity":[93],"Analysis),":[94],"two-level":[96],"framework":[97],"jointly":[99],"optimizes":[100],"calibration-data":[101,112,197],"composition":[102,198],"mixed-precision":[104],"bit":[105,134],"allocation.":[106,135],"Specifically,":[107],"searches":[109],"mixture":[113],"training-free":[116],"gradient-trace":[117],"alignment":[118],"criterion,":[119],"then":[121],"aggregates":[122],"perplexity":[123],"reasoning-oriented":[125],"signals":[127],"to":[128],"guide":[129],"both":[130],"inter-layer":[131],"intra-layer":[133],"Experiments":[136],"LLaMA-3-8B":[138],"Qwen2.5-7B":[140],"precision":[143,160],"inversion:":[144],"appropriately":[145],"allocated":[146],"3.5-bit":[147],"match":[150],"or":[151,166],"surpass":[152],"less":[153],"task-aware":[154],"4-bit":[155,170],"baselines.":[156],"At":[157],"average":[159],"of":[161],"3.5":[162],"bits,":[163],"matches":[165],"outperforms":[167],"several":[168],"competitive":[169],"uniform":[171],"baselines":[172],"aggregate":[174],"accuracy,":[175],"over":[178],"strongest":[180],"W3":[181],"baseline":[182],"GSM8K":[184],"more":[186],"than":[187],"20":[188],"absolute":[189],"points":[190],"LLaMA-3-8B.":[192],"These":[193],"results":[194],"substantially":[199],"affects":[200],"task-sensitive":[201],"quantization,":[202],"factor":[204],"underexplored":[205],"prior":[207],"work.":[208]},"counts_by_year":[],"updated_date":"2026-07-28T07:46:37.118299","created_date":"2026-07-03T00:00:00"}