{"id":"https://openalex.org/W4404355604","doi":"https://doi.org/10.48550/arxiv.2410.15483","title":"Understanding Forgetting in LLM Supervised Fine-Tuning and Preference Learning -- A Convex Optimization Perspective","display_name":"Understanding Forgetting in LLM Supervised Fine-Tuning and Preference Learning -- A Convex Optimization Perspective","publication_year":2024,"publication_date":"2024-10-20","ids":{"openalex":"https://openalex.org/W4404355604","doi":"https://doi.org/10.48550/arxiv.2410.15483"},"language":"en","primary_location":{"id":"pmh:oai:arXiv.org:2410.15483","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2410.15483","pdf_url":"https://arxiv.org/pdf/2410.15483","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/2410.15483","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5046658261","display_name":"Heshan Fernando","orcid":"https://orcid.org/0000-0003-2930-8925","observed_orcids":["https://orcid.org/0000-0003-2930-8925"]},"institutions":[{"id":"https://openalex.org/I165799507","display_name":"Rensselaer Polytechnic Institute","ror":"https://ror.org/01rtyzb94","country_code":"US","type":"education","lineage":["https://openalex.org/I165799507"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Fernando, Heshan","raw_affiliation_strings":["Rensselaer Polytechnic Institute"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Rensselaer Polytechnic Institute","institution_ids":["https://openalex.org/I165799507"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5100359752","display_name":"Han Shen","orcid":"https://orcid.org/0000-0001-6714-5237","observed_orcids":["https://orcid.org/0000-0001-6714-5237","https://orcid.org/0000-0001-9869-2591"]},"institutions":[{"id":"https://openalex.org/I165799507","display_name":"Rensselaer Polytechnic Institute","ror":"https://ror.org/01rtyzb94","country_code":"US","type":"education","lineage":["https://openalex.org/I165799507"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Shen, Han","raw_affiliation_strings":["Rensselaer Polytechnic Institute"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Rensselaer Polytechnic Institute","institution_ids":["https://openalex.org/I165799507"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5089095944","display_name":"Parikshit Ram","orcid":"https://orcid.org/0000-0002-9456-029X","observed_orcids":["https://orcid.org/0000-0002-9456-029X"]},"institutions":[{"id":"https://openalex.org/I1341412227","display_name":"IBM (United States)","ror":"https://ror.org/05hh8d621","country_code":"US","type":"company","lineage":["https://openalex.org/I1341412227"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Ram, Parikshit","raw_affiliation_strings":["IBM Research"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"IBM Research","institution_ids":["https://openalex.org/I1341412227"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5100753053","display_name":"Yi Xin Zhou","orcid":"https://orcid.org/0000-0002-1460-322X","observed_orcids":["https://orcid.org/0000-0002-1460-322X","https://orcid.org/0009-0000-7926-5033","https://orcid.org/0009-0004-7669-4421","https://orcid.org/0000-0002-8520-8227"]},"institutions":[{"id":"https://openalex.org/I1341412227","display_name":"IBM (United States)","ror":"https://ror.org/05hh8d621","country_code":"US","type":"company","lineage":["https://openalex.org/I1341412227"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Zhou, Yi","raw_affiliation_strings":["IBM Research"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"IBM Research","institution_ids":["https://openalex.org/I1341412227"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5035277014","display_name":"Horst Samulowitz","orcid":"https://orcid.org/0000-0002-6780-3217","observed_orcids":["https://orcid.org/0000-0002-6780-3217"]},"institutions":[{"id":"https://openalex.org/I1341412227","display_name":"IBM (United States)","ror":"https://ror.org/05hh8d621","country_code":"US","type":"company","lineage":["https://openalex.org/I1341412227"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Samulowitz, Horst","raw_affiliation_strings":["IBM Research"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"IBM Research","institution_ids":["https://openalex.org/I1341412227"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5010677582","display_name":"Nathalie Baracaldo","orcid":"https://orcid.org/0000-0001-9469-045X","observed_orcids":["https://orcid.org/0000-0001-9469-045X"]},"institutions":[{"id":"https://openalex.org/I1341412227","display_name":"IBM (United States)","ror":"https://ror.org/05hh8d621","country_code":"US","type":"company","lineage":["https://openalex.org/I1341412227"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Baracaldo, Nathalie","raw_affiliation_strings":["IBM Research"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"IBM Research","institution_ids":["https://openalex.org/I1341412227"]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5100783476","display_name":"Tianyi Chen","orcid":"https://orcid.org/0000-0003-3477-1439","observed_orcids":["https://orcid.org/0000-0003-3477-1439","https://orcid.org/0000-0001-8893-3486"]},"institutions":[{"id":"https://openalex.org/I165799507","display_name":"Rensselaer Polytechnic Institute","ror":"https://ror.org/01rtyzb94","country_code":"US","type":"education","lineage":["https://openalex.org/I165799507"]},{"id":"https://openalex.org/I205783295","display_name":"Cornell University","ror":"https://ror.org/05bnh6r87","country_code":"US","type":"education","lineage":["https://openalex.org/I205783295"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Chen, Tianyi","raw_affiliation_strings":["Cornell University","Rensselaer Polytechnic Institute"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Cornell University","institution_ids":["https://openalex.org/I205783295"]},{"raw_affiliation_string":"Rensselaer Polytechnic Institute","institution_ids":["https://openalex.org/I165799507"]}]}],"institutions":[],"countries_distinct_count":1,"institutions_distinct_count":3,"corresponding_author_ids":[],"corresponding_institution_ids":[],"apc_list":null,"apc_paid":null,"fwci":null,"has_fulltext":true,"cited_by_count":1,"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":0.45239999890327454,"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":0.45239999890327454,"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":0.27559998631477356,"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/T11612","display_name":"Stochastic Gradient Optimization Techniques","score":0.03519999980926514,"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/supervised-fine-tuning","display_name":"supervised fine-tuning","score":0.9430000185966492},{"id":"https://openalex.org/keywords/preference-learning","display_name":"preference learning","score":0.9039999842643738},{"id":"https://openalex.org/keywords/reinforcement-learning-from-human-feedback","display_name":"reinforcement learning from human feedback","score":0.9340000152587891},{"id":"https://openalex.org/keywords/dpo","display_name":"DPO","score":0.8119999766349792},{"id":"https://openalex.org/keywords/convex-optimization","display_name":"convex optimization","score":0.7170000076293945},{"id":"https://openalex.org/keywords/convergence-guarantees","display_name":"convergence guarantees","score":0.890999972820282}],"concepts":[{"id":"https://openalex.org/C7149132","wikidata":"https://www.wikidata.org/wiki/Q1377840","display_name":"Forgetting","level":2,"score":0.8627530336380005},{"id":"https://openalex.org/C2781249084","wikidata":"https://www.wikidata.org/wiki/Q908656","display_name":"Preference","level":2,"score":0.7355661392211914},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.4126877784729004},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.3957173824310303},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.3257402181625366},{"id":"https://openalex.org/C180747234","wikidata":"https://www.wikidata.org/wiki/Q23373","display_name":"Cognitive psychology","level":1,"score":0.30799466371536255},{"id":"https://openalex.org/C15744967","wikidata":"https://www.wikidata.org/wiki/Q9418","display_name":"Psychology","level":0,"score":0.30647313594818115},{"id":"https://openalex.org/C162324750","wikidata":"https://www.wikidata.org/wiki/Q8134","display_name":"Economics","level":0,"score":0.17609182000160217},{"id":"https://openalex.org/C175444787","wikidata":"https://www.wikidata.org/wiki/Q39072","display_name":"Microeconomics","level":1,"score":0.0700535774230957}],"mesh":[],"locations_count":2,"locations":[{"id":"pmh:oai:arXiv.org:2410.15483","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2410.15483","pdf_url":"https://arxiv.org/pdf/2410.15483","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.2410.15483","doi":"https://doi.org/10.48550/arxiv.2410.15483","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2410.15483","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":null,"raw_source_name":null,"raw_type":"Preprint"}],"best_oa_location":{"id":"pmh:oai:arXiv.org:2410.15483","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2410.15483","pdf_url":"https://arxiv.org/pdf/2410.15483","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/G1686446952","display_name":"Representation Learning via Variational Mean Field Theory","funder_award_id":"2401297","funder_id":"https://openalex.org/F4320306076","funder_display_name":"National Science Foundation"},{"id":"https://openalex.org/G8382820554","display_name":"Collaborative Research: BLoG: A Bi-Level Optimization Framework for Learning Over Graphs","funder_award_id":"2412486","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"},{"id":"https://openalex.org/F4320309500","display_name":"Rensselaer Polytechnic Institute","ror":"https://ror.org/01rtyzb94"}],"has_content":{"grobid_xml":true,"pdf":true},"content_urls":{"pdf":"https://content.openalex.org/works/W4404355604.pdf","grobid_xml":"https://content.openalex.org/works/W4404355604.grobid-xml"},"referenced_works_count":0,"referenced_works":[],"related_works":["https://openalex.org/W2961085424","https://openalex.org/W4306674287","https://openalex.org/W3046775127","https://openalex.org/W4394896187","https://openalex.org/W3170094116","https://openalex.org/W4386462264","https://openalex.org/W3107602296","https://openalex.org/W4364306694","https://openalex.org/W4312192474","https://openalex.org/W4283697347"],"abstract_inverted_index":{"The":[0,29],"post-training":[1,34,102,108,119],"of":[2,7,51,100],"LLMs,":[3],"which":[4,82,110],"typically":[5],"consists":[6],"the":[8,14,56,61,67,98],"supervised":[9],"fine-tuning":[10],"(SFT)":[11],"stage":[12,17],"and":[13,25,43,53,80,88,103,115],"preference":[15],"learning":[16],"(RLHF":[18],"or":[19],"DPO),":[20],"is":[21,38,47,140],"crucial":[22],"to":[23,39,77,86,123],"effective":[24],"safe":[26],"LLM":[27,57,130],"applications.":[28],"widely":[30],"adopted":[31],"approach":[32],"in":[33,49],"popular":[35],"open-source":[36],"LLMs":[37],"sequentially":[40],"perform":[41],"SFT":[42,52],"RLHF/DPO.":[44],"However,":[45],"this":[46],"suboptimal":[48],"terms":[50],"RLHF/DPO":[54],"trade-off:":[55],"gradually":[58],"forgets":[59],"about":[60],"first":[62],"stage's":[63,69],"training":[64],"when":[65],"undergoing":[66],"second":[68],"training.":[70],"This":[71],"sequential":[72,101,118],"paradigm":[73],"persists":[74],"largely":[75],"due":[76],"its":[78,93],"simplicity":[79],"modularity,":[81],"make":[83],"it":[84],"easier":[85],"implement":[87],"manage":[89],"at":[90,142],"scale":[91],"despite":[92],"limitations.":[94],"We":[95],"theoretically":[96],"prove":[97],"sub-optimality":[99],"propose":[104],"a":[105],"practical":[106],"joint":[107],"framework":[109],"has":[111],"theoretical":[112],"convergence":[113],"guarantees":[114],"empirically":[116],"outperforms":[117],"framework,":[120],"with":[121],"up":[122],"23%":[124],"overall":[125],"performance":[126],"improvement":[127],"across":[128],"multiple":[129],"evaluation":[131],"benchmarks,":[132],"while":[133],"having":[134],"minimal":[135],"computational":[136],"overhead.":[137],"Our":[138],"code":[139],"available":[141],"https://github.com/heshandevaka/XRIGHT.":[143]},"counts_by_year":[{"year":2025,"cited_by_count":1}],"updated_date":"2026-10-16T09:59:22.235173","created_date":"2025-10-10T00:00:00"}