{"id":"https://openalex.org/W7166807563","doi":"https://doi.org/10.18653/v1/2026.acl-long.1976","title":"TLPO: Token-Level Policy Optimization for Mitigating Language Confusion in Large Language Models","display_name":"TLPO: Token-Level Policy Optimization for Mitigating Language Confusion in Large Language Models","publication_year":2026,"publication_date":"2026-01-01","ids":{"openalex":"https://openalex.org/W7166807563","doi":"https://doi.org/10.18653/v1/2026.acl-long.1976"},"language":"en","primary_location":{"id":"doi:10.18653/v1/2026.acl-long.1976","doi":"https://doi.org/10.18653/v1/2026.acl-long.1976","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.acl-long.1976","pdf_url":null,"source":{"id":"https://openalex.org/S4393915740","display_name":"Proceedings of the conference - Association for Computational Linguistics. Meeting","issn_l":"0736-587X","issn":["0736-587X"],"is_oa":false,"is_in_doaj":false,"is_core":false,"listed_in":["norway-1"],"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":"Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)","raw_type":"proceedings-article"},"volume":null,"type":"conference-paper","is_traditional":true,"indexed_in":["arxiv","crossref","datacite"],"open_access":{"is_oa":true,"oa_status":"hybrid","oa_url":"https://doi.org/10.18653/v1/2026.acl-long.1976","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5134881012","display_name":"Jinho Choo","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Jinho Choo","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5134894773","display_name":"JunSeung Lee","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"JunSeung Lee","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5037195485","display_name":"Jimyeong Kim","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Jimyeong Kim","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5006136942","display_name":"Yeeho Song","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Yeeho Song","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5127984624","display_name":"S. K. Hong","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"S. K. Hong","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5062265168","display_name":"Yeong\u2010Dae Kwon","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Yeong-Dae Kwon","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]}],"institutions":[],"countries_distinct_count":0,"institutions_distinct_count":0,"corresponding_author_ids":[],"corresponding_institution_ids":[],"apc_list":null,"apc_paid":null,"fwci":0.0,"has_fulltext":true,"cited_by_count":0,"citation_normalized_percentile":{"value":0.34919028,"is_in_top_1_percent":false,"is_in_top_10_percent":false},"cited_by_percentile_year":null,"biblio":{"volume":null,"issue":null,"first_page":"42670","last_page":"42690"},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T10181","display_name":"Natural Language Processing Techniques","score":0.5396000146865845,"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/T10181","display_name":"Natural Language Processing Techniques","score":0.5396000146865845,"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":0.3917999863624573,"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/T12031","display_name":"Speech and dialogue systems","score":0.01730000041425228,"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/language-confusion","display_name":"language confusion","score":0.9679999947547913},{"id":"https://openalex.org/keywords/large-language-models","display_name":"large language models","score":0.8479999899864197},{"id":"https://openalex.org/keywords/multilingual-llms","display_name":"multilingual LLMs","score":0.7820000052452087},{"id":"https://openalex.org/keywords/language-consistency","display_name":"language consistency","score":0.824999988079071},{"id":"https://openalex.org/keywords/dpo","display_name":"DPO","score":0.5720000267028809},{"id":"https://openalex.org/keywords/group-relative-policy-optimization","display_name":"Group Relative Policy Optimization","score":0.968999981880188}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.6373000144958496},{"id":"https://openalex.org/C2781140086","wikidata":"https://www.wikidata.org/wiki/Q557945","display_name":"Confusion","level":2,"score":0.4203000068664551},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.39309999346733093},{"id":"https://openalex.org/C137293760","wikidata":"https://www.wikidata.org/wiki/Q3621696","display_name":"Language model","level":2,"score":0.36730000376701355},{"id":"https://openalex.org/C204321447","wikidata":"https://www.wikidata.org/wiki/Q30642","display_name":"Natural language processing","level":1,"score":0.36329999566078186},{"id":"https://openalex.org/C195324797","wikidata":"https://www.wikidata.org/wiki/Q33742","display_name":"Natural language","level":2,"score":0.349700003862381},{"id":"https://openalex.org/C129792486","wikidata":"https://www.wikidata.org/wiki/Q1050419","display_name":"Language identification","level":3,"score":0.3418000042438507},{"id":"https://openalex.org/C41895202","wikidata":"https://www.wikidata.org/wiki/Q8162","display_name":"Linguistics","level":1,"score":0.3230000138282776},{"id":"https://openalex.org/C179603123","wikidata":"https://www.wikidata.org/wiki/Q1941921","display_name":"Modeling language","level":3,"score":0.28870001435279846},{"id":"https://openalex.org/C2779313563","wikidata":"https://www.wikidata.org/wiki/Q17072565","display_name":"On Language","level":2,"score":0.27250000834465027},{"id":"https://openalex.org/C136197465","wikidata":"https://www.wikidata.org/wiki/Q1729295","display_name":"Variety (cybernetics)","level":2,"score":0.26820001006126404},{"id":"https://openalex.org/C2777152325","wikidata":"https://www.wikidata.org/wiki/Q108163","display_name":"Proposition","level":2,"score":0.2587999999523163},{"id":"https://openalex.org/C2779530757","wikidata":"https://www.wikidata.org/wiki/Q1207505","display_name":"Quality (philosophy)","level":2,"score":0.25769999623298645},{"id":"https://openalex.org/C26517878","wikidata":"https://www.wikidata.org/wiki/Q228039","display_name":"Key (lock)","level":2,"score":0.25690001249313354},{"id":"https://openalex.org/C155092808","wikidata":"https://www.wikidata.org/wiki/Q182557","display_name":"Computational linguistics","level":2,"score":0.25589999556541443},{"id":"https://openalex.org/C116834253","wikidata":"https://www.wikidata.org/wiki/Q2039217","display_name":"Identification (biology)","level":2,"score":0.2500999867916107}],"mesh":[],"locations_count":3,"locations":[{"id":"doi:10.18653/v1/2026.acl-long.1976","doi":"https://doi.org/10.18653/v1/2026.acl-long.1976","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.acl-long.1976","pdf_url":null,"source":{"id":"https://openalex.org/S4393915740","display_name":"Proceedings of the conference - Association for Computational Linguistics. Meeting","issn_l":"0736-587X","issn":["0736-587X"],"is_oa":false,"is_in_doaj":false,"is_core":false,"listed_in":["norway-1"],"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":"Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)","raw_type":"proceedings-article"},{"id":"pmh:oai:arXiv.org:2604.26553","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2604.26553","pdf_url":"https://arxiv.org/pdf/2604.26553","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":"other-oa","license_id":"https://openalex.org/licenses/other-oa","version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},{"id":"doi:10.48550/arxiv.2604.26553","doi":"https://doi.org/10.48550/arxiv.2604.26553","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2604.26553","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":"doi:10.18653/v1/2026.acl-long.1976","doi":"https://doi.org/10.18653/v1/2026.acl-long.1976","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.acl-long.1976","pdf_url":null,"source":{"id":"https://openalex.org/S4393915740","display_name":"Proceedings of the conference - Association for Computational Linguistics. Meeting","issn_l":"0736-587X","issn":["0736-587X"],"is_oa":false,"is_in_doaj":false,"is_core":false,"listed_in":["norway-1"],"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":"Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)","raw_type":"proceedings-article"},"sustainable_development_goals":[],"study_designs":[],"awards":[],"funders":[],"has_content":{"grobid_xml":true,"pdf":true},"content_urls":{"pdf":"https://content.openalex.org/works/W7166807563.pdf","grobid_xml":"https://content.openalex.org/works/W7166807563.grobid-xml"},"referenced_works_count":14,"referenced_works":["https://openalex.org/W2736601468","https://openalex.org/W2794325560","https://openalex.org/W3134642945","https://openalex.org/W3210904070","https://openalex.org/W4308760226","https://openalex.org/W4327810158","https://openalex.org/W4378465252","https://openalex.org/W4390723197","https://openalex.org/W4391555989","https://openalex.org/W4391631327","https://openalex.org/W4401306886","https://openalex.org/W4414281281","https://openalex.org/W4414939355","https://openalex.org/W6947991961"],"related_works":["https://openalex.org/W7166824454","https://openalex.org/W7162694115","https://openalex.org/W7166851230","https://openalex.org/W7024992033","https://openalex.org/W4402670143","https://openalex.org/W3027602624","https://openalex.org/W4411119523","https://openalex.org/W4415795638","https://openalex.org/W7196948905","https://openalex.org/W4412945699"],"abstract_inverted_index":{"Large":[0],"language":[1,24,78,115,139],"models":[2],"(LLMs)":[3],"demonstrate":[4,131],"strong":[5],"multilingual":[6,126],"capabilities,":[7,55],"yet":[8],"often":[9],"fail":[10],"to":[11,49,76,100],"consistently":[12],"generate":[13],"responses":[14,45],"in":[15,137],"the":[16,41,57,94,119],"intended":[17],"language,":[18],"exhibiting":[19],"a":[20,72,97,105],"phenomenon":[21],"known":[22],"as":[23,34],"confusion.":[25],"Prior":[26],"mitigation":[27,113],"approaches":[28],"based":[29],"on":[30,124],"sequence-level":[31],"fine-tuning,":[32],"such":[33],"DPO,":[35],"ORPO,":[36],"and":[37,46,92],"GRPO,":[38],"operate":[39],"at":[40,104],"level":[42],"of":[43,52,114],"entire":[44],"can":[47],"lead":[48],"unintended":[50],"degradation":[51],"general":[53,121],"model":[54],"motivating":[56],"need":[58],"for":[59],"more":[60],"fine-grained":[61],"alternatives.":[62],"To":[63],"address":[64],"this,":[65],"we":[66],"introduce":[67],"Token-Level":[68],"Policy":[69],"Optimization":[70],"(TLPO),":[71],"fine-tuning":[73],"framework":[74],"designed":[75],"mitigate":[77],"confusion":[79,116],"through":[80],"localized,":[81],"token-level":[82],"updates.":[83],"TLPO":[84,133],"identifies":[85],"error-prone":[86],"positions,":[87],"explores":[88],"alternative":[89],"candidate":[90],"tokens,":[91],"updates":[93],"policy":[95],"using":[96],"tailored":[98],"objective":[99],"suppress":[101],"error-inducing":[102],"outputs":[103],"granular":[106],"level.":[107],"This":[108],"selective":[109],"intervention":[110],"enables":[111],"effective":[112],"without":[117],"compromising":[118],"model's":[120],"abilities.":[122],"Experiments":[123],"multiple":[125],"LLMs":[127],"across":[128],"diverse":[129],"languages":[130],"that":[132],"significantly":[134],"outperforms":[135],"baselines":[136],"improving":[138],"consistency":[140],"while":[141],"preserving":[142],"downstream":[143],"task":[144],"accuracy.":[145]},"counts_by_year":[],"updated_date":"2026-11-08T12:12:28.789561","created_date":"2026-07-02T00:00:00"}