{"id":"https://openalex.org/W4386907142","doi":"https://doi.org/10.48550/arxiv.2309.10105","title":"Understanding Catastrophic Forgetting in Language Models via Implicit Inference","display_name":"Understanding Catastrophic Forgetting in Language Models via Implicit Inference","publication_year":2023,"publication_date":"2023-09-18","ids":{"openalex":"https://openalex.org/W4386907142","doi":"https://doi.org/10.48550/arxiv.2309.10105"},"language":"en","primary_location":{"id":"pmh:oai:arXiv.org:2309.10105","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2309.10105","pdf_url":"https://arxiv.org/pdf/2309.10105","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":"cc-by","license_id":"https://openalex.org/licenses/cc-by","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/2309.10105","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5077282083","display_name":"Suhas Kotha","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I74973139","display_name":"Carnegie Mellon University","ror":"https://ror.org/05x2bcf33","country_code":"US","type":"education","lineage":["https://openalex.org/I74973139"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Kotha, Suhas","raw_affiliation_strings":["Carnegie Mellon University"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Carnegie Mellon University","institution_ids":["https://openalex.org/I74973139"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5035773495","display_name":"Jacob M. Springer","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I74973139","display_name":"Carnegie Mellon University","ror":"https://ror.org/05x2bcf33","country_code":"US","type":"education","lineage":["https://openalex.org/I74973139"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Springer, Jacob Mitchell","raw_affiliation_strings":["Carnegie Mellon University"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Carnegie Mellon University","institution_ids":["https://openalex.org/I74973139"]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5031731960","display_name":"Aditi Raghunathan","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I74973139","display_name":"Carnegie Mellon University","ror":"https://ror.org/05x2bcf33","country_code":"US","type":"education","lineage":["https://openalex.org/I74973139"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Raghunathan, Aditi","raw_affiliation_strings":["Carnegie Mellon University"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Carnegie Mellon University","institution_ids":["https://openalex.org/I74973139"]}]}],"institutions":[],"countries_distinct_count":1,"institutions_distinct_count":1,"corresponding_author_ids":[],"corresponding_institution_ids":["https://openalex.org/I74973139"],"apc_list":null,"apc_paid":null,"fwci":null,"has_fulltext":true,"cited_by_count":6,"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/T11307","display_name":"Domain Adaptation and Few-Shot Learning","score":0.61080002784729,"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/T11307","display_name":"Domain Adaptation and Few-Shot Learning","score":0.61080002784729,"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":0.30790001153945923,"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":0.037700001150369644,"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/catastrophic-forgetting","display_name":"catastrophic forgetting","score":0.9990000128746033},{"id":"https://openalex.org/keywords/instruction-tuning","display_name":"instruction tuning","score":0.6439999938011169},{"id":"https://openalex.org/keywords/reinforcement-learning-from-human-feedback","display_name":"reinforcement learning from human feedback","score":0.9139999747276306},{"id":"https://openalex.org/keywords/harmful-content-generation","display_name":"harmful content generation","score":0.5249999761581421},{"id":"https://openalex.org/keywords/in-context-learning","display_name":"in-context learning","score":0.7749999761581421},{"id":"https://openalex.org/keywords/language-models","display_name":"language models","score":0.4350000023841858}],"concepts":[{"id":"https://openalex.org/C157524613","wikidata":"https://www.wikidata.org/wiki/Q2828883","display_name":"Fine-tuning","level":2,"score":0.8213472962379456},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.7507810592651367},{"id":"https://openalex.org/C7149132","wikidata":"https://www.wikidata.org/wiki/Q1377840","display_name":"Forgetting","level":2,"score":0.6909496188163757},{"id":"https://openalex.org/C2780451532","wikidata":"https://www.wikidata.org/wiki/Q759676","display_name":"Task (project management)","level":2,"score":0.6692134737968445},{"id":"https://openalex.org/C2776214188","wikidata":"https://www.wikidata.org/wiki/Q408386","display_name":"Inference","level":2,"score":0.6259605288505554},{"id":"https://openalex.org/C98045186","wikidata":"https://www.wikidata.org/wiki/Q205663","display_name":"Process (computing)","level":2,"score":0.5565205812454224},{"id":"https://openalex.org/C2779343474","wikidata":"https://www.wikidata.org/wiki/Q3109175","display_name":"Context (archaeology)","level":2,"score":0.5300426483154297},{"id":"https://openalex.org/C97541855","wikidata":"https://www.wikidata.org/wiki/Q830687","display_name":"Reinforcement learning","level":2,"score":0.48237577080726624},{"id":"https://openalex.org/C137293760","wikidata":"https://www.wikidata.org/wiki/Q3621696","display_name":"Language model","level":2,"score":0.4530666470527649},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.4144773483276367},{"id":"https://openalex.org/C180747234","wikidata":"https://www.wikidata.org/wiki/Q23373","display_name":"Cognitive psychology","level":1,"score":0.1591757833957672},{"id":"https://openalex.org/C15744967","wikidata":"https://www.wikidata.org/wiki/Q9418","display_name":"Psychology","level":0,"score":0.08888402581214905},{"id":"https://openalex.org/C199360897","wikidata":"https://www.wikidata.org/wiki/Q9143","display_name":"Programming language","level":1,"score":0.07845214009284973},{"id":"https://openalex.org/C162324750","wikidata":"https://www.wikidata.org/wiki/Q8134","display_name":"Economics","level":0,"score":0.0},{"id":"https://openalex.org/C62520636","wikidata":"https://www.wikidata.org/wiki/Q944","display_name":"Quantum mechanics","level":1,"score":0.0},{"id":"https://openalex.org/C86803240","wikidata":"https://www.wikidata.org/wiki/Q420","display_name":"Biology","level":0,"score":0.0},{"id":"https://openalex.org/C151730666","wikidata":"https://www.wikidata.org/wiki/Q7205","display_name":"Paleontology","level":1,"score":0.0},{"id":"https://openalex.org/C121332964","wikidata":"https://www.wikidata.org/wiki/Q413","display_name":"Physics","level":0,"score":0.0},{"id":"https://openalex.org/C187736073","wikidata":"https://www.wikidata.org/wiki/Q2920921","display_name":"Management","level":1,"score":0.0}],"mesh":[],"locations_count":2,"locations":[{"id":"pmh:oai:arXiv.org:2309.10105","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2309.10105","pdf_url":"https://arxiv.org/pdf/2309.10105","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":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},{"id":"doi:10.48550/arxiv.2309.10105","doi":"https://doi.org/10.48550/arxiv.2309.10105","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2309.10105","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":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":null,"is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"Preprint"}],"best_oa_location":{"id":"pmh:oai:arXiv.org:2309.10105","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2309.10105","pdf_url":"https://arxiv.org/pdf/2309.10105","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":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},"sustainable_development_goals":[],"sustainable_development_goals_aurora":[],"study_designs":[],"awards":[{"id":"https://openalex.org/G8051438088","display_name":"Graduate Research Fellowship Program (GRFP)","funder_award_id":"2140739","funder_id":"https://openalex.org/F4320306076","funder_display_name":"National Science Foundation"}],"funders":[{"id":"https://openalex.org/F4320306076","display_name":"National Science Foundation","ror":"https://ror.org/021nxhr62"}],"has_content":{"pdf":true,"grobid_xml":true},"content_urls":{"pdf":"https://content.openalex.org/works/W4386907142.pdf","grobid_xml":"https://content.openalex.org/works/W4386907142.grobid-xml"},"referenced_works_count":0,"referenced_works":[],"related_works":["https://openalex.org/W7163174650","https://openalex.org/W4417308166","https://openalex.org/W4381573334","https://openalex.org/W4385573164","https://openalex.org/W7134626658","https://openalex.org/W4392089972","https://openalex.org/W4416009630","https://openalex.org/W4416037226","https://openalex.org/W4404356317","https://openalex.org/W7101792598"],"abstract_inverted_index":{"We":[0,54],"lack":[1],"a":[2,30],"systematic":[3],"understanding":[4],"of":[5,8,49,63,108],"the":[6,25,41,47,61,64,75,88,93,98,109,136],"effects":[7],"fine-tuning":[9,27,42,68,76,94,118,169],"(via":[10],"methods":[11],"such":[12],"as":[13],"instruction-tuning":[14],"or":[15],"reinforcement":[16],"learning":[17,147],"from":[18,92],"human":[19],"feedback),":[20],"particularly":[21],"on":[22,38,51],"tasks":[23,39,73],"outside":[24],"narrow":[26],"distribution.":[28,77],"In":[29],"simplified":[31],"scenario,":[32],"we":[33,81,102,123],"demonstrate":[34],"that":[35,56,67,104],"improving":[36],"performance":[37],"within":[40],"data":[43],"distribution":[44,95],"comes":[45],"at":[46],"expense":[48],"capabilities":[50,111,130],"other":[52],"tasks.":[53],"hypothesize":[55],"language":[57],"models":[58],"implicitly":[59],"infer":[60],"task":[62,89],"prompt":[65],"and":[66,101],"skews":[69],"this":[70,105],"inference":[71],"towards":[72],"in":[74,112,131,170],"To":[78],"test":[79],"this,":[80],"propose":[82],"Conjugate":[83],"Prompting,":[84],"which":[85],"artificially":[86],"makes":[87],"look":[90],"farther":[91],"while":[96],"requiring":[97],"same":[99],"capability,":[100],"find":[103],"recovers":[106],"some":[107],"pretraining":[110],"our":[113],"synthetic":[114],"setup.":[115],"Since":[116],"real-world":[117],"distributions":[119],"are":[120],"predominantly":[121],"English,":[122],"apply":[124],"conjugate":[125],"prompting":[126],"to":[127,138,144],"recover":[128,145],"pretrained":[129],"LLMs":[132],"by":[133,167],"simply":[134],"translating":[135],"prompts":[137],"different":[139],"languages.":[140],"This":[141],"allows":[142],"us":[143],"in-context":[146],"abilities":[148],"lost":[149,156],"via":[150],"instruction":[151],"tuning,":[152],"natural":[153],"reasoning":[154],"capability":[155],"during":[157],"code":[158],"fine-tuning,":[159],"and,":[160],"more":[161],"concerningly,":[162],"harmful":[163],"content":[164],"generation":[165],"suppressed":[166],"safety":[168],"chatbots":[171],"like":[172],"ChatGPT.":[173]},"counts_by_year":[{"year":2026,"cited_by_count":1},{"year":2025,"cited_by_count":2},{"year":2024,"cited_by_count":3}],"updated_date":"2026-10-16T09:59:22.235173","created_date":"2023-09-21T00:00:00"}