{"id":"https://openalex.org/W7164041871","doi":"https://doi.org/10.48550/arxiv.2606.09012","title":"Understanding Quantization-Aware Training: Gradients at Quantized Weights Bias to the Low-Loss Basin","display_name":"Understanding Quantization-Aware Training: Gradients at Quantized Weights Bias to the Low-Loss Basin","publication_year":2026,"publication_date":"2026-06-08","ids":{"openalex":"https://openalex.org/W7164041871","doi":"https://doi.org/10.48550/arxiv.2606.09012"},"language":null,"primary_location":{"id":"doi:10.48550/arxiv.2606.09012","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2606.09012","pdf_url":null,"source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":null,"issn":null,"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":null,"license_id":null,"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.2606.09012","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5138245142","display_name":"Hanyang Li","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Li, Hanyang","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5102224245","display_name":"Jianhao Ma","orcid":"https://orcid.org/0000-0003-1440-6645"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Ma, Jianhao","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5138207093","display_name":"Ying Cui","orcid":"https://orcid.org/0009-0000-8274-9821"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Cui, Ying","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/T10036","display_name":"Advanced Neural Network Applications","score":0.5239999890327454,"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":0.5239999890327454,"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/T11307","display_name":"Domain Adaptation and Few-Shot Learning","score":0.2484000027179718,"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/T11689","display_name":"Adversarial Robustness in Machine Learning","score":0.06530000269412994,"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/quantization","display_name":"Quantization (signal processing)","score":0.7749000191688538},{"id":"https://openalex.org/keywords/iterated-function","display_name":"Iterated function","score":0.5254999995231628},{"id":"https://openalex.org/keywords/rounding","display_name":"Rounding","score":0.47049999237060547},{"id":"https://openalex.org/keywords/spurious-relationship","display_name":"Spurious relationship","score":0.3709000051021576},{"id":"https://openalex.org/keywords/grid","display_name":"Grid","score":0.32089999318122864}],"concepts":[{"id":"https://openalex.org/C28855332","wikidata":"https://www.wikidata.org/wiki/Q198099","display_name":"Quantization (signal processing)","level":2,"score":0.7749000191688538},{"id":"https://openalex.org/C33923547","wikidata":"https://www.wikidata.org/wiki/Q395","display_name":"Mathematics","level":0,"score":0.5738000273704529},{"id":"https://openalex.org/C140479938","wikidata":"https://www.wikidata.org/wiki/Q5254619","display_name":"Iterated function","level":2,"score":0.5254999995231628},{"id":"https://openalex.org/C136625980","wikidata":"https://www.wikidata.org/wiki/Q663208","display_name":"Rounding","level":2,"score":0.47049999237060547},{"id":"https://openalex.org/C11413529","wikidata":"https://www.wikidata.org/wiki/Q8366","display_name":"Algorithm","level":1,"score":0.4025999903678894},{"id":"https://openalex.org/C97256817","wikidata":"https://www.wikidata.org/wiki/Q1462316","display_name":"Spurious relationship","level":2,"score":0.3709000051021576},{"id":"https://openalex.org/C187691185","wikidata":"https://www.wikidata.org/wiki/Q2020720","display_name":"Grid","level":2,"score":0.32089999318122864},{"id":"https://openalex.org/C109007969","wikidata":"https://www.wikidata.org/wiki/Q749565","display_name":"Structural basin","level":2,"score":0.3199999928474426},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.29159998893737793},{"id":"https://openalex.org/C28719098","wikidata":"https://www.wikidata.org/wiki/Q44946","display_name":"Point (geometry)","level":2,"score":0.2824999988079071},{"id":"https://openalex.org/C143397304","wikidata":"https://www.wikidata.org/wiki/Q630426","display_name":"Geometric quantization","level":5,"score":0.27000001072883606},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.26969999074935913},{"id":"https://openalex.org/C28826006","wikidata":"https://www.wikidata.org/wiki/Q33521","display_name":"Applied mathematics","level":1,"score":0.26649999618530273}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.48550/arxiv.2606.09012","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2606.09012","pdf_url":null,"source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":null,"issn":null,"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":null,"license_id":null,"version":null,"is_accepted":false,"is_published":null,"raw_source_name":null,"raw_type":"Preprint"}],"best_oa_location":{"id":"doi:10.48550/arxiv.2606.09012","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2606.09012","pdf_url":null,"source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":null,"issn":null,"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":null,"license_id":null,"version":null,"is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"Preprint"},"sustainable_development_goals":[],"awards":[],"funders":[],"has_content":{"pdf":false,"grobid_xml":false},"content_urls":null,"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Post-training":[0],"quantization":[1,19,101,212],"(PTQ)":[2],"converts":[3],"a":[4,54,72,76,79,86,95,120,141,184,189],"trained":[5],"full-precision":[6,68,155],"model":[7,67],"into":[8,20,176],"low-bit":[9],"weights":[10,151],"without":[11],"task-level":[12],"retraining,":[13],"while":[14,90,152],"quantization-aware":[15],"training":[16,22,69],"(QAT)":[17],"incorporates":[18],"the":[21,49,83,100,106,126,148,158,162,177,215,222],"loop.":[23],"Although":[24],"PTQ":[25,61,110,191,220],"is":[26,42,103],"efficient":[27],"and":[28,63,114,165,194,206,221],"often":[29,47],"accurate":[30],"at":[31,38,147],"moderate":[32],"bitwidths,":[33],"it":[34,144],"can":[35,46,118],"fail":[36],"sharply":[37],"aggressive":[39],"bitwidths;":[40],"QAT":[41,64,139,197],"more":[43],"expensive":[44],"but":[45],"recover":[48],"lost":[50],"accuracy.":[51],"We":[52,66,179],"propose":[53],"unified":[55],"geometric":[56,190],"framework":[57],"that":[58,170],"explains":[59],"both":[60],"failure":[62,192,218],"recovery.":[65],"as":[70],"following":[71],"low-loss":[73,131],"\\emph{river}":[74],"inside":[75],"wider":[77],"\\emph{valley}:":[78],"normal":[80],"neighborhood":[81],"of":[82,219,226],"river":[84],"forms":[85],"nearly":[87],"flat":[88],"\\emph{basin},":[89],"leaving":[91],"this":[92,136,181],"basin":[93,107,127],"incurs":[94],"sharp":[96],"loss":[97],"increase.":[98],"When":[99],"grid":[102],"comparable":[104],"to":[105,160],"width,":[108],"local":[109,185,200],"objectives,":[111],"including":[112],"rounding":[113],"Hessian-based":[115],"second-order":[116],"reconstruction,":[117],"select":[119],"high-loss":[121],"deployed":[122,149],"quantized":[123,132,150,173],"point":[124],"outside":[125],"even":[128],"when":[129],"nearby":[130],"points":[133],"exist.":[134],"In":[135],"regime,":[137],"straight-through-estimator-based":[138],"has":[140],"useful":[142],"bias:":[143],"evaluates":[145],"gradients":[146],"updating":[153],"latent":[154],"weights,":[156],"causing":[157],"gradient":[159],"sense":[161],"valley":[163],"wall":[164],"acquire":[166],"an":[167],"inward":[168],"component":[169],"steers":[171],"subsequent":[172],"iterates":[174],"back":[175],"basin.":[178],"formalize":[180],"mechanism":[182,225],"through":[183],"landscape":[186],"model,":[187],"construct":[188],"mode,":[193],"prove":[195],"finite-time":[196],"recovery":[198,224],"under":[199,209],"quantizer-compatibility":[201],"assumptions.":[202],"Experiments":[203],"across":[204],"vision":[205],"language":[207],"models":[208],"multiple":[210],"neural-network":[211],"schemes":[213],"corroborate":[214],"predicted":[216],"basin-crossing":[217],"corresponding":[223],"QAT.":[227]},"counts_by_year":[],"updated_date":"2026-07-01T06:00:48.157686","created_date":"2026-06-10T00:00:00"}