{"id":"https://openalex.org/W4223891676","doi":"https://doi.org/10.48550/arxiv.2204.05832","title":"What Language Model Architecture and Pretraining Objective Work Best for Zero-Shot Generalization?","display_name":"What Language Model Architecture and Pretraining Objective Work Best for Zero-Shot Generalization?","publication_year":2022,"publication_date":"2022-04-12","ids":{"openalex":"https://openalex.org/W4223891676","doi":"https://doi.org/10.48550/arxiv.2204.05832"},"language":"en","primary_location":{"id":"pmh:oai:arXiv.org:2204.05832","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2204.05832","pdf_url":"https://arxiv.org/pdf/2204.05832","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/2204.05832","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":null,"display_name":"Wang, Thomas","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Wang, Thomas","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5052454696","display_name":"Adam Roberts","orcid":"https://orcid.org/0000-0003-1621-1964","observed_orcids":["https://orcid.org/0000-0003-1621-1964"]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Roberts, Adam","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5012892331","display_name":"Daniel Hesslow","orcid":"https://orcid.org/0000-0002-7279-7042","observed_orcids":["https://orcid.org/0000-0002-7279-7042"]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Hesslow, Daniel","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5084957527","display_name":"Teven Le Scao","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Scao, Teven Le","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5051828575","display_name":"Hyung Won Chung","orcid":"https://orcid.org/0000-0002-1280-9953","observed_orcids":["https://orcid.org/0000-0002-1280-9953"]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Chung, Hyung Won","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5090038537","display_name":"Iz Beltagy","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I4210156221","display_name":"Allen Institute for Artificial Intelligence","ror":"https://ror.org/05w520734","country_code":"US","type":"nonprofit","lineage":["https://openalex.org/I4210156221"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Beltagy, Iz","raw_affiliation_strings":["Allen Institute for AI"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Allen Institute for AI","institution_ids":["https://openalex.org/I4210156221"]}]},{"author_position":"middle","author":{"id":null,"display_name":"Launay, Julien","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I4210103330","display_name":"\u00c9cole Normale Sup\u00e9rieure","ror":"https://ror.org/01dp7jr64","country_code":"BI","type":"education","lineage":["https://openalex.org/I4210103330"]}],"countries":["BI"],"is_corresponding":false,"raw_author_name":"Launay, Julien","raw_affiliation_strings":["LPENS , \u00c9cole Normale Sup\u00e9rieure"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"LPENS , \u00c9cole Normale Sup\u00e9rieure","institution_ids":["https://openalex.org/I4210103330"]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5045077843","display_name":"Colin Raffel","orcid":"https://orcid.org/0000-0002-5009-8048","observed_orcids":["https://orcid.org/0000-0002-5009-8048"]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Raffel, Colin","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]}],"institutions":[],"countries_distinct_count":2,"institutions_distinct_count":2,"corresponding_author_ids":[],"corresponding_institution_ids":[],"apc_list":null,"apc_paid":null,"fwci":null,"has_fulltext":true,"cited_by_count":23,"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/T10028","display_name":"Topic Modeling","score":0.9667999744415283,"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/T10028","display_name":"Topic Modeling","score":0.9667999744415283,"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/T10181","display_name":"Natural Language Processing Techniques","score":0.01979999989271164,"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":0.009999999776482582,"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/zero-shot-generalization","display_name":"zero-shot generalization","score":0.9959999918937683},{"id":"https://openalex.org/keywords/pre-trained-language-models","display_name":"pre-trained language models","score":0.6420000195503235},{"id":"https://openalex.org/keywords/transformer","display_name":"transformer models","score":0.7590000033378601},{"id":"https://openalex.org/keywords/encoder-decoder-architecture","display_name":"encoder-decoder architecture","score":0.9520000219345093},{"id":"https://openalex.org/keywords/masked-language-modeling","display_name":"masked language modeling","score":0.9990000128746033},{"id":"https://openalex.org/keywords/text-to-text-models","display_name":"text-to-text models","score":0.7770000100135803},{"id":"https://openalex.org/keywords/unsupervised-pretraining","display_name":"unsupervised pre-training","score":0.7419999837875366},{"id":"https://openalex.org/keywords/model-adaptation","display_name":"model adaptation","score":0.36399999260902405}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.8078539371490479},{"id":"https://openalex.org/C137293760","wikidata":"https://www.wikidata.org/wiki/Q3621696","display_name":"Language model","level":2,"score":0.7423654794692993},{"id":"https://openalex.org/C177148314","wikidata":"https://www.wikidata.org/wiki/Q170084","display_name":"Generalization","level":2,"score":0.625632643699646},{"id":"https://openalex.org/C159877910","wikidata":"https://www.wikidata.org/wiki/Q2202883","display_name":"Autoregressive model","level":2,"score":0.5817387104034424},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.5323989987373352},{"id":"https://openalex.org/C11671645","wikidata":"https://www.wikidata.org/wiki/Q5054567","display_name":"Causal model","level":2,"score":0.5181494951248169},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.49883079528808594},{"id":"https://openalex.org/C2780451532","wikidata":"https://www.wikidata.org/wiki/Q759676","display_name":"Task (project management)","level":2,"score":0.432644784450531},{"id":"https://openalex.org/C33923547","wikidata":"https://www.wikidata.org/wiki/Q395","display_name":"Mathematics","level":0,"score":0.08264485001564026},{"id":"https://openalex.org/C134306372","wikidata":"https://www.wikidata.org/wiki/Q7754","display_name":"Mathematical analysis","level":1,"score":0.0},{"id":"https://openalex.org/C105795698","wikidata":"https://www.wikidata.org/wiki/Q12483","display_name":"Statistics","level":1,"score":0.0},{"id":"https://openalex.org/C187736073","wikidata":"https://www.wikidata.org/wiki/Q2920921","display_name":"Management","level":1,"score":0.0},{"id":"https://openalex.org/C162324750","wikidata":"https://www.wikidata.org/wiki/Q8134","display_name":"Economics","level":0,"score":0.0},{"id":"https://openalex.org/C149782125","wikidata":"https://www.wikidata.org/wiki/Q160039","display_name":"Econometrics","level":1,"score":0.0}],"mesh":[],"locations_count":2,"locations":[{"id":"pmh:oai:arXiv.org:2204.05832","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2204.05832","pdf_url":"https://arxiv.org/pdf/2204.05832","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.2204.05832","doi":"https://doi.org/10.48550/arxiv.2204.05832","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2204.05832","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:2204.05832","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2204.05832","pdf_url":"https://arxiv.org/pdf/2204.05832","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":[{"id":"https://openalex.org/F4320322892","display_name":"Centre National de la Recherche Scientifique","ror":"https://ror.org/02feahw73"},{"id":"https://openalex.org/F4320326256","display_name":"Grand \u00c9quipement National De Calcul Intensif","ror":"https://ror.org/0010d1q40"}],"has_content":{"pdf":false,"grobid_xml":false},"content_urls":null,"referenced_works_count":0,"referenced_works":[],"related_works":["https://openalex.org/W4385571231","https://openalex.org/W3205068155","https://openalex.org/W4385567008","https://openalex.org/W4411119955","https://openalex.org/W4385571291","https://openalex.org/W4385572426","https://openalex.org/W3198002980","https://openalex.org/W3194309076","https://openalex.org/W4385571701","https://openalex.org/W4320235125"],"abstract_inverted_index":{"Large":[0],"pretrained":[1,184,193,218],"Transformer":[2],"language":[3,93,141,165,208],"models":[4,37,73,105,136,154,185,196,221],"have":[5],"been":[6,43],"shown":[7],"to":[8,126],"exhibit":[9,144],"zero-shot":[10,65,147],"generalization,":[11],"i.e.":[12],"they":[13,22],"can":[14,197,222],"perform":[15,172],"a":[16,55,163,211],"wide":[17],"variety":[18],"of":[19,47,58,183],"tasks":[20],"that":[21,121,133,192,217],"were":[23],"not":[24],"explicitly":[25],"trained":[26,84,137,161],"on.":[27],"However,":[28,153],"the":[29,119,145,173,181],"architectures":[30,79,187],"and":[31,40,61,74,82,91,95,98,188,238],"pretraining":[32,88],"objectives":[33,89],"used":[34],"across":[35,186],"state-of-the-art":[36],"differ":[38],"significantly,":[39],"there":[41],"has":[42],"limited":[44],"systematic":[45],"comparison":[46],"these":[48],"factors.":[49],"In":[50,67],"this":[51],"work,":[52],"we":[53,69,215],"present":[54],"large-scale":[56],"evaluation":[57],"modeling":[59,142,166,209],"choices":[60],"their":[62,159],"impact":[63],"on":[64,71,138,158],"generalization.":[66],"particular,":[68],"focus":[70],"text-to-text":[72],"experiment":[75],"with":[76,85,97,106,155,162],"three":[77],"model":[78],"(causal/non-causal":[80],"decoder-only":[81,135],"encoder-decoder),":[83],"two":[86],"different":[87],"(autoregressive":[90],"masked":[92,164],"modeling),":[94],"evaluated":[96],"without":[99],"multitask":[100,170,235],"prompted":[101],"finetuning.":[102,236],"We":[103,178,190],"train":[104],"over":[107],"5":[108],"billion":[109,115],"parameters":[110],"for":[111],"more":[112],"than":[113],"170":[114],"tokens,":[116],"thereby":[117],"increasing":[118],"likelihood":[120],"our":[122,176],"conclusions":[123],"will":[124],"transfer":[125],"even":[127],"larger":[128],"scales.":[129],"Our":[130],"experiments":[131],"show":[132],"causal":[134,203,219],"an":[139],"autoregressive":[140,207],"objective":[143,167],"strongest":[146],"generalization":[148],"after":[149,234],"purely":[150],"unsupervised":[151],"pretraining.":[152],"non-causal":[156,194,227],"visibility":[157],"input":[160],"followed":[168],"by":[169],"finetuning":[171],"best":[174],"among":[175],"experiments.":[177],"therefore":[179],"consider":[180],"adaptation":[182],"objectives.":[189],"find":[191,216],"decoder":[195,204,220,228],"be":[198,223],"adapted":[199,225],"into":[200,226],"performant":[201],"generative":[202],"models,":[205,229],"using":[206],"as":[210],"downstream":[212],"task.":[213],"Furthermore,":[214],"efficiently":[224],"ultimately":[230],"achieving":[231],"competitive":[232],"performance":[233],"Code":[237],"checkpoints":[239],"are":[240],"available":[241],"at":[242],"https://github.com/bigscience-workshop/architecture-objective.":[243]},"counts_by_year":[{"year":2025,"cited_by_count":3},{"year":2024,"cited_by_count":4},{"year":2023,"cited_by_count":10},{"year":2022,"cited_by_count":6}],"updated_date":"2026-10-16T09:59:22.235173","created_date":"2022-04-19T00:00:00"}