{"id":"https://openalex.org/W7214036683","doi":"https://doi.org/10.48550/arxiv.2609.23585","title":"Global Ranks Survive, Selected Heads Shift: BOS-Sink Topology under 4-bit Weight-Only Quantization","display_name":"Global Ranks Survive, Selected Heads Shift: BOS-Sink Topology under 4-bit Weight-Only Quantization","publication_year":2026,"publication_date":"2026-09-20","ids":{"openalex":"https://openalex.org/W7214036683","doi":"https://doi.org/10.48550/arxiv.2609.23585"},"language":"en","primary_location":{"id":"pmh:oai:arXiv.org:2609.23585","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2609.23585","pdf_url":"https://arxiv.org/pdf/2609.23585","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":null,"license_id":null,"version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},"volume":null,"type":"preprint","indexed_in":["arxiv","datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://arxiv.org/pdf/2609.23585","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5152481494","display_name":"Kuanlin Chen","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I25846049","display_name":"National Tsing Hua University","ror":"https://ror.org/00zdnkx70","country_code":"TW","type":"education","lineage":["https://openalex.org/I25846049"]}],"countries":["TW"],"is_corresponding":false,"raw_author_name":"Chen, Kuanlin","raw_affiliation_strings":["Independent Researcher Taoyuan , Taiwan","National Tsing Hua University Hsinchu , Taiwan"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Independent Researcher Taoyuan , Taiwan","institution_ids":[]},{"raw_affiliation_string":"National Tsing Hua University Hsinchu , Taiwan","institution_ids":["https://openalex.org/I25846049"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5152477807","display_name":"Chen-Wei Kuo","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I25846049","display_name":"National Tsing Hua University","ror":"https://ror.org/00zdnkx70","country_code":"TW","type":"education","lineage":["https://openalex.org/I25846049"]}],"countries":["TW"],"is_corresponding":false,"raw_author_name":"Kuo, Chen-Wei","raw_affiliation_strings":["National Tsing Hua University Hsinchu , Taiwan"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"National Tsing Hua University Hsinchu , Taiwan","institution_ids":["https://openalex.org/I25846049"]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5152475777","display_name":"Cheng-En Ou","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Ou, Cheng-En","raw_affiliation_strings":["Independent Researcher Taoyuan , Taiwan"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Independent Researcher Taoyuan , Taiwan","institution_ids":[]}]}],"institutions":[],"countries_distinct_count":1,"institutions_distinct_count":1,"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":0.9023000001907349,"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":0.9023000001907349,"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":0.0142000000923872,"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":0.005499999970197678,"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/weight-only-quantization","display_name":"weight-only quantization","score":0.7929999828338623},{"id":"https://openalex.org/keywords/post-training-quantization","display_name":"post-training quantization","score":0.8769999742507935},{"id":"https://openalex.org/keywords/recalibration","display_name":"recalibration","score":0.5400000214576721},{"id":"https://openalex.org/keywords/edge-deployment","display_name":"edge deployment","score":0.628000020980835},{"id":"https://openalex.org/keywords/jetson-orin-nx","display_name":"Jetson Orin NX","score":0.8610000014305115},{"id":"https://openalex.org/keywords/qwen","display_name":"Qwen","score":0.7080000042915344},{"id":"https://openalex.org/keywords/llama-3-2","display_name":"Llama 3.2","score":0.921999990940094}],"concepts":[{"id":"https://openalex.org/C184720557","wikidata":"https://www.wikidata.org/wiki/Q7825049","display_name":"Topology (electrical circuits)","level":2,"score":0.5831000208854675},{"id":"https://openalex.org/C28855332","wikidata":"https://www.wikidata.org/wiki/Q198099","display_name":"Quantization (signal processing)","level":2,"score":0.49630001187324524},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.4830999970436096},{"id":"https://openalex.org/C33923547","wikidata":"https://www.wikidata.org/wiki/Q395","display_name":"Mathematics","level":0,"score":0.4625000059604645},{"id":"https://openalex.org/C42812","wikidata":"https://www.wikidata.org/wiki/Q1082910","display_name":"Partition (number theory)","level":2,"score":0.45879998803138733},{"id":"https://openalex.org/C11413529","wikidata":"https://www.wikidata.org/wiki/Q8366","display_name":"Algorithm","level":1,"score":0.4564000070095062},{"id":"https://openalex.org/C206588197","wikidata":"https://www.wikidata.org/wiki/Q846574","display_name":"Reuse","level":2,"score":0.41019999980926514},{"id":"https://openalex.org/C89128539","wikidata":"https://www.wikidata.org/wiki/Q1949963","display_name":"Statistic","level":2,"score":0.38830000162124634},{"id":"https://openalex.org/C176217482","wikidata":"https://www.wikidata.org/wiki/Q860554","display_name":"Metric (unit)","level":2,"score":0.3790999948978424},{"id":"https://openalex.org/C45340560","wikidata":"https://www.wikidata.org/wiki/Q215382","display_name":"Disjoint sets","level":2,"score":0.3750999867916107},{"id":"https://openalex.org/C203519979","wikidata":"https://www.wikidata.org/wiki/Q865360","display_name":"Jaccard index","level":3,"score":0.3668999969959259},{"id":"https://openalex.org/C112972136","wikidata":"https://www.wikidata.org/wiki/Q7595718","display_name":"Stability (learning theory)","level":2,"score":0.3377000093460083},{"id":"https://openalex.org/C177264268","wikidata":"https://www.wikidata.org/wiki/Q1514741","display_name":"Set (abstract data type)","level":2,"score":0.31220000982284546},{"id":"https://openalex.org/C143050476","wikidata":"https://www.wikidata.org/wiki/Q194502","display_name":"Sink (geography)","level":2,"score":0.3001999855041504},{"id":"https://openalex.org/C2776436953","wikidata":"https://www.wikidata.org/wiki/Q5163215","display_name":"Consistency (knowledge bases)","level":2,"score":0.289900004863739},{"id":"https://openalex.org/C105339364","wikidata":"https://www.wikidata.org/wiki/Q2297740","display_name":"Software deployment","level":2,"score":0.2872999906539917},{"id":"https://openalex.org/C199845137","wikidata":"https://www.wikidata.org/wiki/Q145490","display_name":"Network topology","level":2,"score":0.2797999978065491},{"id":"https://openalex.org/C165838908","wikidata":"https://www.wikidata.org/wiki/Q736777","display_name":"Calibration","level":2,"score":0.27810001373291016},{"id":"https://openalex.org/C132459708","wikidata":"https://www.wikidata.org/wiki/Q744069","display_name":"Extrapolation","level":2,"score":0.27639999985694885},{"id":"https://openalex.org/C188198153","wikidata":"https://www.wikidata.org/wiki/Q1613840","display_name":"Limiting","level":2,"score":0.27399998903274536},{"id":"https://openalex.org/C53533937","wikidata":"https://www.wikidata.org/wiki/Q185020","display_name":"Histogram","level":3,"score":0.258899986743927},{"id":"https://openalex.org/C164660894","wikidata":"https://www.wikidata.org/wiki/Q2037833","display_name":"Piecewise","level":2,"score":0.25679999589920044},{"id":"https://openalex.org/C157497606","wikidata":"https://www.wikidata.org/wiki/Q603227","display_name":"traceroute","level":3,"score":0.2558000087738037},{"id":"https://openalex.org/C2778476105","wikidata":"https://www.wikidata.org/wiki/Q628539","display_name":"Workload","level":2,"score":0.251800000667572}],"mesh":[],"locations_count":2,"locations":[{"id":"pmh:oai:arXiv.org:2609.23585","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2609.23585","pdf_url":"https://arxiv.org/pdf/2609.23585","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":null,"license_id":null,"version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},{"id":"doi:10.48550/arxiv.2609.23585","doi":"https://doi.org/10.48550/arxiv.2609.23585","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2609.23585","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":null,"license_id":null,"version":null,"is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"Preprint"}],"best_oa_location":{"id":"pmh:oai:arXiv.org:2609.23585","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2609.23585","pdf_url":"https://arxiv.org/pdf/2609.23585","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":null,"license_id":null,"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/W7214036683.pdf","grobid_xml":"https://content.openalex.org/works/W7214036683.grobid-xml"},"referenced_works_count":0,"referenced_works":[],"related_works":["https://openalex.org/W7214733366","https://openalex.org/W7213402146","https://openalex.org/W7133333059","https://openalex.org/W7167618119","https://openalex.org/W7213957151","https://openalex.org/W7151249066","https://openalex.org/W7083835173","https://openalex.org/W7134960504","https://openalex.org/W7164955587","https://openalex.org/W7214108747"],"abstract_inverted_index":{"Sink-aware":[0],"deployment":[1],"may":[2],"identify":[3],"important":[4],"first-token":[5],"attention":[6],"heads":[7],"before":[8],"a":[9,110,136,156],"model":[10,100],"is":[11,25,78,196],"quantized,":[12],"then":[13],"reuse":[14],"that":[15],"map":[16],"at":[17,68,140],"the":[18,118,141,160,171,179,185],"edge.":[19],"We":[20],"test":[21],"when":[22],"this":[23],"shortcut":[24],"safe":[26],"for":[27,122,128,145,151,159,184],"4-bit":[28,131],"NF4":[29],"weight-only":[30],"post-training":[31],"quantization":[32],"(PTQ).":[33],"Our":[34],"Sink":[35],"Topology":[36],"Consistency":[37],"(STC)":[38],"metrics":[39],"separate":[40],"global":[41,63,87,198],"rank":[42],"preservation,":[43],"top-$k$":[44,75],"set":[45],"overlap,":[46],"and":[47,51,61,149,208],"layerwise":[48],"sink-mass":[49],"shift,":[50,112],"distinguish":[52],"per-input":[53],"sensitivity":[54],"from":[55],"calibration-map":[56],"transfer.":[57],"Across":[58],"Qwen2.5-0.5B,":[59],"Qwen2.5-1.5B,":[60],"Llama-3.2-1B,":[62,152],"bf16-to-4-bit":[64],"ranks":[65],"remain":[66],"high":[67],"4,096":[69],"tokens":[70],"($\u03c1_s":[71],"\\geq":[72],"0.980$),":[73],"yet":[74],"Jaccard":[76],"overlap":[77,114],"only":[79,165],"0.619-0.793,":[80],"corresponding":[81],"to":[82],"76.5-88.5%":[83],"membership":[84],"retention.":[85],"The":[86,193],"statistic":[88],"also":[89],"masks":[90],"local":[91],"failures:":[92],"terminal":[93],"Qwen":[94,124,147,162],"layers":[95,167],"shift":[96],"by":[97],"6.2-7.9x":[98],"their":[99],"means,":[101],"whereas":[102],"Llama-3.2-1B":[103],"shows":[104],"low,":[105],"nearly":[106],"uniform":[107],"drift.":[108],"Under":[109],"C4-to-LongBench":[111],"cross-domain":[113,209],"degrades":[115],"more":[116],"than":[117],"within-domain":[119],"precision":[120],"comparison":[121],"both":[123,146],"models,":[125,163],"but":[126,202],"not":[127,154,169],"Llama-3.2-1B.":[129],"Matched-domain":[130],"recalibration":[132],"reaches":[133],"90%":[134],"of":[135],"split-half":[137],"stability":[138,173],"plateau":[139],"smallest":[142],"tested":[143],"$n=8$":[144],"models":[148,187],"$n=32$":[150],"though":[153],"as":[155],"sharp":[157],"threshold;":[158],"two":[161,186],"updating":[164],"selected":[166],"does":[168],"reach":[170],"full-map":[172],"criterion.":[174],"On":[175],"Jetson":[176],"Orin":[177],"NX,":[178],"16-sample":[180],"workload":[181],"takes":[182],"seconds":[183],"with":[188],"valid":[189],"on-device":[190],"sink":[191],"measurements.":[192],"practical":[194],"message":[195],"precise:":[197],"rankings":[199],"often":[200],"transfer,":[201],"discrete":[203],"head":[204],"sets,":[205],"layer-local":[206],"policies,":[207],"calibration":[210],"should":[211],"be":[212],"revalidated":[213],"after":[214],"quantization.":[215]},"counts_by_year":[],"updated_date":"2026-10-28T10:33:24.094632","created_date":"2026-09-23T00:00:00"}