{"id":"https://openalex.org/W7166804156","doi":"https://doi.org/10.18653/v1/2026.findings-acl.1668","title":"Bias Dynamics in BabyLMs: Towards a Compute-Efficient Sandbox for Democratising Pre-Training Debiasing","display_name":"Bias Dynamics in BabyLMs: Towards a Compute-Efficient Sandbox for Democratising Pre-Training Debiasing","publication_year":2026,"publication_date":"2026-01-01","ids":{"openalex":"https://openalex.org/W7166804156","doi":"https://doi.org/10.18653/v1/2026.findings-acl.1668"},"language":"en","primary_location":{"id":"doi:10.18653/v1/2026.findings-acl.1668","doi":"https://doi.org/10.18653/v1/2026.findings-acl.1668","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.findings-acl.1668","pdf_url":"https://aclanthology.org/2026.findings-acl.1668.pdf","source":{"id":"https://openalex.org/S4363605144","display_name":"Findings of the Association for Computational Linguistics: ACL (Findings of ACL)","issn_l":null,"issn":null,"is_oa":false,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":null,"host_organization_name":null,"host_organization_lineage":[],"host_organization_lineage_names":[],"type":"conference"},"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Findings of the Association for Computational Linguistics: ACL 2026","raw_type":"proceedings-article"},"volume":null,"type":"conference-paper","indexed_in":["crossref"],"open_access":{"is_oa":true,"oa_status":"hybrid","oa_url":"https://aclanthology.org/2026.findings-acl.1668.pdf","any_repository_has_fulltext":false},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5073848053","display_name":"Filip Trhl\u00edk","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I241749","display_name":"University of Cambridge","ror":"https://ror.org/013meh722","country_code":"GB","type":"education","lineage":["https://openalex.org/I241749"]}],"countries":["GB"],"is_corresponding":false,"raw_author_name":"Filip Trhl\u00edk","raw_affiliation_strings":["Department of Computer Science & Technology , University of Cambridge , U.K. ALTA Institute , University of Cambridge , U.K"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Department of Computer Science & Technology , University of Cambridge , U.K. ALTA Institute , University of Cambridge , U.K","institution_ids":["https://openalex.org/I241749"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5069029343","display_name":"Andrew Caines","orcid":"https://orcid.org/0000-0001-9647-4902","observed_orcids":["https://orcid.org/0000-0001-9647-4902"]},"institutions":[{"id":"https://openalex.org/I241749","display_name":"University of Cambridge","ror":"https://ror.org/013meh722","country_code":"GB","type":"education","lineage":["https://openalex.org/I241749"]}],"countries":["GB"],"is_corresponding":false,"raw_author_name":"Andrew Caines","raw_affiliation_strings":["Department of Computer Science & Technology , University of Cambridge , U.K. ALTA Institute , University of Cambridge , U.K"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Department of Computer Science & Technology , University of Cambridge , U.K. ALTA Institute , University of Cambridge , U.K","institution_ids":["https://openalex.org/I241749"]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5013999263","display_name":"Paula J. Buttery","orcid":"https://orcid.org/0000-0003-3874-0656","observed_orcids":["https://orcid.org/0000-0003-3874-0656"]},"institutions":[{"id":"https://openalex.org/I241749","display_name":"University of Cambridge","ror":"https://ror.org/013meh722","country_code":"GB","type":"education","lineage":["https://openalex.org/I241749"]}],"countries":["GB"],"is_corresponding":false,"raw_author_name":"Paula Buttery","raw_affiliation_strings":["Department of Computer Science & Technology , University of Cambridge , U.K. ALTA Institute , University of Cambridge , U.K"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Department of Computer Science & Technology , University of Cambridge , U.K. ALTA Institute , University of Cambridge , U.K","institution_ids":["https://openalex.org/I241749"]}]}],"institutions":[],"countries_distinct_count":1,"institutions_distinct_count":1,"corresponding_author_ids":[],"corresponding_institution_ids":["https://openalex.org/I241749"],"apc_list":null,"apc_paid":null,"fwci":0.0,"has_fulltext":true,"cited_by_count":0,"citation_normalized_percentile":{"value":0.41180706,"is_in_top_1_percent":false,"is_in_top_10_percent":false},"cited_by_percentile_year":null,"biblio":{"volume":null,"issue":null,"first_page":"33356","last_page":"33377"},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T10883","display_name":"Ethics and Social Impacts of AI","score":0.8773999810218811,"subfield":{"id":"https://openalex.org/subfields/3311","display_name":"Safety Research"},"field":{"id":"https://openalex.org/fields/33","display_name":"Social Sciences"},"domain":{"id":"https://openalex.org/domains/2","display_name":"Social Sciences"}},"topics":[{"id":"https://openalex.org/T10883","display_name":"Ethics and Social Impacts of AI","score":0.8773999810218811,"subfield":{"id":"https://openalex.org/subfields/3311","display_name":"Safety Research"},"field":{"id":"https://openalex.org/fields/33","display_name":"Social Sciences"},"domain":{"id":"https://openalex.org/domains/2","display_name":"Social Sciences"}},{"id":"https://openalex.org/T12262","display_name":"Hate Speech and Cyberbullying Detection","score":0.04690000042319298,"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":0.033900000154972076,"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/babylm","display_name":"BabyLM","score":0.7590000033378601},{"id":"https://openalex.org/keywords/small-language-models","display_name":"small language models","score":0.6179999709129333},{"id":"https://openalex.org/keywords/bert","display_name":"BERT","score":0.6320000290870667},{"id":"https://openalex.org/keywords/gender-imbalance","display_name":"gender imbalance","score":0.6069999933242798},{"id":"https://openalex.org/keywords/toxicity","display_name":"toxicity","score":0.7680000066757202},{"id":"https://openalex.org/keywords/bias-mitigation","display_name":"bias mitigation","score":0.453000009059906},{"id":"https://openalex.org/keywords/computational-efficiency","display_name":"computational efficiency","score":0.4580000042915344}],"concepts":[{"id":"https://openalex.org/C2779458634","wikidata":"https://www.wikidata.org/wiki/Q24963715","display_name":"Debiasing","level":2,"score":0.7843999862670898},{"id":"https://openalex.org/C167981075","wikidata":"https://www.wikidata.org/wiki/Q2667186","display_name":"Sandbox (software development)","level":2,"score":0.6409000158309937},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.47600001096725464},{"id":"https://openalex.org/C145912823","wikidata":"https://www.wikidata.org/wiki/Q113558","display_name":"Dynamics (music)","level":2,"score":0.47519999742507935},{"id":"https://openalex.org/C111472728","wikidata":"https://www.wikidata.org/wiki/Q9471","display_name":"Epistemology","level":1,"score":0.3537999987602234},{"id":"https://openalex.org/C15744967","wikidata":"https://www.wikidata.org/wiki/Q9418","display_name":"Psychology","level":0,"score":0.30550000071525574},{"id":"https://openalex.org/C127627568","wikidata":"https://www.wikidata.org/wiki/Q1639361","display_name":"Sociotechnical system","level":2,"score":0.30410000681877136},{"id":"https://openalex.org/C77805123","wikidata":"https://www.wikidata.org/wiki/Q161272","display_name":"Social psychology","level":1,"score":0.28349998593330383},{"id":"https://openalex.org/C2779343474","wikidata":"https://www.wikidata.org/wiki/Q3109175","display_name":"Context (archaeology)","level":2,"score":0.2784999907016754},{"id":"https://openalex.org/C144024400","wikidata":"https://www.wikidata.org/wiki/Q21201","display_name":"Sociology","level":0,"score":0.27129998803138733}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.18653/v1/2026.findings-acl.1668","doi":"https://doi.org/10.18653/v1/2026.findings-acl.1668","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.findings-acl.1668","pdf_url":"https://aclanthology.org/2026.findings-acl.1668.pdf","source":{"id":"https://openalex.org/S4363605144","display_name":"Findings of the Association for Computational Linguistics: ACL (Findings of ACL)","issn_l":null,"issn":null,"is_oa":false,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":null,"host_organization_name":null,"host_organization_lineage":[],"host_organization_lineage_names":[],"type":"conference"},"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Findings of the Association for Computational Linguistics: ACL 2026","raw_type":"proceedings-article"}],"best_oa_location":{"id":"doi:10.18653/v1/2026.findings-acl.1668","doi":"https://doi.org/10.18653/v1/2026.findings-acl.1668","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.findings-acl.1668","pdf_url":"https://aclanthology.org/2026.findings-acl.1668.pdf","source":{"id":"https://openalex.org/S4363605144","display_name":"Findings of the Association for Computational Linguistics: ACL (Findings of ACL)","issn_l":null,"issn":null,"is_oa":false,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":null,"host_organization_name":null,"host_organization_lineage":[],"host_organization_lineage_names":[],"type":"conference"},"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Findings of the Association for Computational Linguistics: ACL 2026","raw_type":"proceedings-article"},"sustainable_development_goals":[],"sustainable_development_goals_aurora":[],"study_designs":[],"awards":[],"funders":[{"id":"https://openalex.org/F4320320273","display_name":"University of Cambridge","ror":"https://ror.org/013meh722"}],"has_content":{"pdf":true,"grobid_xml":true},"content_urls":{"pdf":"https://content.openalex.org/works/W7166804156.pdf","grobid_xml":"https://content.openalex.org/works/W7166804156.grobid-xml"},"referenced_works_count":0,"referenced_works":[],"related_works":["https://openalex.org/W4379958573","https://openalex.org/W4285192297","https://openalex.org/W4401043228","https://openalex.org/W7172384673","https://openalex.org/W4402670064","https://openalex.org/W4390963114","https://openalex.org/W7045443536","https://openalex.org/W4285273243","https://openalex.org/W6925379443","https://openalex.org/W6976603073"],"abstract_inverted_index":{"Pre-trained":[0],"language":[1],"models":[2,76],"(LMs)":[3],"have,":[4],"over":[5,175],"the":[6,52,148],"last":[7],"few":[8],"years,":[9],"grown":[10],"substantially":[11],"in":[12,21,26],"both":[13],"societal":[14],"adoption":[15],"and":[16,28,80,88,105,121,127,143,153,190],"training":[17],"costs.This":[18],"rapid":[19],"growth":[20],"size":[22],"has":[23,40],"constrained":[24],"progress":[25],"understanding":[27],"mitigating":[29],"their":[30,114],"biases.Since":[31],"re-training":[32],"LMs":[33],"is":[34],"prohibitively":[35],"expensive,":[36],"most":[37],"debiasing":[38,64,129,136,188],"work":[39],"focused":[41],"on":[42,78,155],"posthoc":[43],"or":[44],"masking-based":[45],"strategies,":[46],"which":[47],"often":[48],"fail":[49],"to":[50,61,109,178,185],"address":[51],"underlying":[53],"causes":[54],"of":[55,91,101,150,196],"bias.In":[56],"this":[57],"work,":[58],"we":[59,71,133],"seek":[60],"democratise":[62,186],"pre-model":[63,135,187],"research":[65,189],"by":[66],"using":[67],"lowcost":[68],"proxy":[69],"models.Specifically,":[70],"investigate":[72],"BabyLMs,":[73,139],"compact":[74],"BERT-like":[75],"trained":[77],"small":[79],"mutable":[81],"corpora":[82],"that":[83,95,160],"can":[84,162],"approximate":[85],"bias":[86,103,156],"acquisition":[87],"learning":[89],"dynamics":[90],"larger":[92],"models.We":[93],"show":[94],"BabyLMs":[96,120,161],"display":[97],"closely":[98],"aligned":[99],"patterns":[100],"intrinsic":[102],"formation":[104],"performance":[106],"development":[107],"compared":[108],"standard":[110],"BERT":[111,122],"models,":[112],"despite":[113],"drastically":[115],"reduced":[116],"size.Furthermore,":[117],"correlations":[118],"between":[119],"hold":[123],"across":[124],"multiple":[125],"intra-model":[126],"post-model":[128],"methods.Leveraging":[130],"these":[131],"similarities,":[132],"conduct":[134],"experiments":[137],"with":[138],"replicating":[140],"prior":[141],"findings":[142],"presenting":[144],"new":[145],"insights":[146],"regarding":[147],"influence":[149],"gender":[151],"imbalance":[152],"toxicity":[154],"formation.Our":[157],"results":[158],"demonstrate":[159],"serve":[163],"as":[164],"an":[165],"effective":[166],"sandbox":[167],"for":[168,198],"large-scale":[169],"LMs,":[170],"reducing":[171],"pre-training":[172],"costs":[173],"from":[174],"500":[176],"GPU-hours":[177],"under":[179],"30":[180],"GPU-hours.This":[181],"provides":[182],"a":[183],"way":[184],"enables":[191],"faster,":[192],"more":[193],"accessible":[194],"exploration":[195],"methods":[197],"building":[199],"fairer":[200],"LMs.":[201]},"counts_by_year":[],"updated_date":"2026-10-28T10:33:24.094632","created_date":"2026-07-02T00:00:00"}