{"id":"https://openalex.org/W7133339030","doi":"https://doi.org/10.48550/arxiv.2603.02170","title":"SageBwd: A Trainable Low-bit Attention","display_name":"SageBwd: A Trainable Low-bit Attention","publication_year":2026,"publication_date":"2026-03-02","ids":{"openalex":"https://openalex.org/W7133339030","doi":"https://doi.org/10.48550/arxiv.2603.02170"},"language":null,"primary_location":{"id":"doi:10.48550/arxiv.2603.02170","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2603.02170","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":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":null,"is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"article"},"type":"preprint","indexed_in":["datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://doi.org/10.48550/arxiv.2603.02170","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5127990449","display_name":"Jintao Zhang","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Zhang, Jintao","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5048011252","display_name":"Marco Tianyu Chen","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Chen, Marco","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5128008037","display_name":"Haoxu Wang","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Wang, Haoxu","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5128011958","display_name":"Kai Jiang","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Jiang, Kai","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5128032005","display_name":"Ion Stoica","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Stoica, Ion","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5127893828","display_name":"Joseph E. Gonzalez","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Gonzalez, Joseph E.","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5128034983","display_name":"Jianfei Chen","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Chen, Jianfei","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5128022739","display_name":"Jun Zhu","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Zhu, Jun","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]}],"institutions":[],"countries_distinct_count":0,"institutions_distinct_count":8,"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":0.40450000762939453,"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":0.40450000762939453,"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/T11689","display_name":"Adversarial Robustness in Machine Learning","score":0.21629999577999115,"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":0.13079999387264252,"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":0.5223000049591064},{"id":"https://openalex.org/keywords/training","display_name":"Training (meteorology)","score":0.3287999927997589},{"id":"https://openalex.org/keywords/training-set","display_name":"Training set","score":0.2976999878883362},{"id":"https://openalex.org/keywords/matrix","display_name":"Matrix (chemical analysis)","score":0.29440000653266907},{"id":"https://openalex.org/keywords/pattern-recognition","display_name":"Pattern recognition (psychology)","score":0.29010000824928284}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.6883000135421753},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.5940999984741211},{"id":"https://openalex.org/C28855332","wikidata":"https://www.wikidata.org/wiki/Q198099","display_name":"Quantization (signal processing)","level":2,"score":0.5223000049591064},{"id":"https://openalex.org/C2777211547","wikidata":"https://www.wikidata.org/wiki/Q17141490","display_name":"Training (meteorology)","level":2,"score":0.3287999927997589},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.3221000134944916},{"id":"https://openalex.org/C51632099","wikidata":"https://www.wikidata.org/wiki/Q3985153","display_name":"Training set","level":2,"score":0.2976999878883362},{"id":"https://openalex.org/C106487976","wikidata":"https://www.wikidata.org/wiki/Q685816","display_name":"Matrix (chemical analysis)","level":2,"score":0.29440000653266907},{"id":"https://openalex.org/C153180895","wikidata":"https://www.wikidata.org/wiki/Q7148389","display_name":"Pattern recognition (psychology)","level":2,"score":0.29010000824928284},{"id":"https://openalex.org/C31972630","wikidata":"https://www.wikidata.org/wiki/Q844240","display_name":"Computer vision","level":1,"score":0.28439998626708984},{"id":"https://openalex.org/C28490314","wikidata":"https://www.wikidata.org/wiki/Q189436","display_name":"Speech recognition","level":1,"score":0.27379998564720154},{"id":"https://openalex.org/C2779343474","wikidata":"https://www.wikidata.org/wiki/Q3109175","display_name":"Context (archaeology)","level":2,"score":0.2621000111103058},{"id":"https://openalex.org/C11413529","wikidata":"https://www.wikidata.org/wiki/Q8366","display_name":"Algorithm","level":1,"score":0.26010000705718994},{"id":"https://openalex.org/C99498987","wikidata":"https://www.wikidata.org/wiki/Q2210247","display_name":"Noise (video)","level":3,"score":0.25119999051094055}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.48550/arxiv.2603.02170","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2603.02170","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":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":null,"is_accepted":false,"is_published":null,"raw_source_name":null,"raw_type":"article"}],"best_oa_location":{"id":"doi:10.48550/arxiv.2603.02170","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2603.02170","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":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":null,"is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"article"},"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":{"Low-bit":[0],"attention,":[1],"such":[2],"as":[3,7],"SageAttention,":[4],"has":[5],"emerged":[6],"an":[8],"effective":[9],"approach":[10],"for":[11,93,130],"accelerating":[12],"model":[13],"inference,":[14],"but":[15],"its":[16],"applicability":[17],"to":[18,52,119],"training":[19,95,131],"remains":[20,128],"poorly":[21],"understood.":[22],"In":[23,58],"prior":[24],"work,":[25,60],"we":[26,61,81],"introduced":[27],"SageBwd,":[28],"a":[29,48,83],"trainable":[30],"INT8":[31],"attention":[32,38,54,73],"that":[33,69],"quantizes":[34],"six":[35],"of":[36],"seven":[37],"matrix":[39],"multiplications":[40],"while":[41,133],"preserving":[42],"fine-tuning":[43],"performance.":[44],"However,":[45],"SageBwd":[46,70,118],"exhibited":[47],"persistent":[49],"performance":[50,122],"gap":[51,65],"full-precision":[53,72],"(FPA)":[55],"during":[56,74,138],"pre-training.":[57,139],"this":[59,64],"investigate":[62],"why":[63],"occurs":[66],"and":[67,78,87,125],"demonstrate":[68],"matches":[71],"pretraining.":[75],"Through":[76],"experiments":[77],"theoretical":[79],"analysis,":[80],"reach":[82],"few":[84],"important":[85],"insights":[86],"conclusions:":[88],"(i)":[89],"QK-norm":[90],"is":[91],"necessary":[92],"stable":[94],"at":[96],"large":[97],"tokens":[98,114],"per":[99,115],"step,":[100],"(ii)":[101],"quantization":[102],"errors":[103],"primarily":[104],"arise":[105],"from":[106],"the":[107],"backward-pass":[108],"score":[109],"gradient":[110],"dS,":[111],"(iii)":[112],"reducing":[113],"step":[116],"enables":[117],"match":[120],"FPA":[121],"in":[123],"pre-training,":[124],"(iv)":[126],"K-smoothing":[127],"essential":[129],"stability,":[132],"Q-smoothing":[134],"provides":[135],"limited":[136],"benefit":[137]},"counts_by_year":[],"updated_date":"2026-06-11T09:08:48.828518","created_date":"2026-03-04T00:00:00"}