{"id":"https://openalex.org/W7166841842","doi":"https://doi.org/10.18653/v1/2026.acl-long.405","title":"Efficient KL Divergence Estimation via Truncated Top-K Integration for Large Language Models","display_name":"Efficient KL Divergence Estimation via Truncated Top-K Integration for Large Language Models","publication_year":2026,"publication_date":"2026-01-01","ids":{"openalex":"https://openalex.org/W7166841842","doi":"https://doi.org/10.18653/v1/2026.acl-long.405"},"language":null,"primary_location":{"id":"doi:10.18653/v1/2026.acl-long.405","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.acl-long.405","pdf_url":"https://aclanthology.org/2026.acl-long.405.pdf","source":null,"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"},"type":"conference-paper","indexed_in":["crossref"],"open_access":{"is_oa":true,"oa_status":"gold","oa_url":"https://aclanthology.org/2026.acl-long.405.pdf","any_repository_has_fulltext":null},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5139784381","display_name":"Xinyuan Wang","orcid":null},"institutions":[{"id":"https://openalex.org/I24943067","display_name":"Fudan University","ror":"https://ror.org/013q1eq08","country_code":"CN","type":"education","lineage":["https://openalex.org/I24943067"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Xinyuan Wang","raw_affiliation_strings":["Fudan University"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Fudan University","institution_ids":["https://openalex.org/I24943067"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5139792978","display_name":"Luozhijie Jin","orcid":null},"institutions":[{"id":"https://openalex.org/I24943067","display_name":"Fudan University","ror":"https://ror.org/013q1eq08","country_code":"CN","type":"education","lineage":["https://openalex.org/I24943067"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Luozhijie Jin","raw_affiliation_strings":["Fudan University"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Fudan University","institution_ids":["https://openalex.org/I24943067"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5139808603","display_name":"Bo Wang","orcid":null},"institutions":[{"id":"https://openalex.org/I24943067","display_name":"Fudan University","ror":"https://ror.org/013q1eq08","country_code":"CN","type":"education","lineage":["https://openalex.org/I24943067"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Bo Wang","raw_affiliation_strings":["Fudan University"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Fudan University","institution_ids":["https://openalex.org/I24943067"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5139828064","display_name":"Yuan Li","orcid":null},"institutions":[{"id":"https://openalex.org/I24943067","display_name":"Fudan University","ror":"https://ror.org/013q1eq08","country_code":"CN","type":"education","lineage":["https://openalex.org/I24943067"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Yuan Li","raw_affiliation_strings":["Fudan University"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Fudan University","institution_ids":["https://openalex.org/I24943067"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5139787321","display_name":"Zhangyue Yin","orcid":null},"institutions":[{"id":"https://openalex.org/I24943067","display_name":"Fudan University","ror":"https://ror.org/013q1eq08","country_code":"CN","type":"education","lineage":["https://openalex.org/I24943067"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Zhangyue Yin","raw_affiliation_strings":["Fudan University"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Fudan University","institution_ids":["https://openalex.org/I24943067"]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5139804213","display_name":"Xipeng Qiu","orcid":null},"institutions":[{"id":"https://openalex.org/I24943067","display_name":"Fudan University","ror":"https://ror.org/013q1eq08","country_code":"CN","type":"education","lineage":["https://openalex.org/I24943067"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Xipeng Qiu","raw_affiliation_strings":["Fudan University"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Fudan University","institution_ids":["https://openalex.org/I24943067"]}]}],"institutions":[],"countries_distinct_count":1,"institutions_distinct_count":1,"corresponding_author_ids":[],"corresponding_institution_ids":["https://openalex.org/I24943067"],"apc_list":null,"apc_paid":null,"fwci":0.0,"has_fulltext":true,"cited_by_count":0,"citation_normalized_percentile":{"value":0.83165762,"is_in_top_1_percent":false,"is_in_top_10_percent":false},"cited_by_percentile_year":null,"biblio":{"volume":null,"issue":null,"first_page":"8973","last_page":"8985"},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T10181","display_name":"Natural Language Processing Techniques","score":0.27320000529289246,"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":0.27320000529289246,"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/T10201","display_name":"Speech Recognition and Synthesis","score":0.22840000689029694,"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":0.16850000619888306,"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/divergence","display_name":"Divergence (linguistics)","score":0.45190000534057617},{"id":"https://openalex.org/keywords/kullback\u2013leibler-divergence","display_name":"Kullback\u2013Leibler divergence","score":0.44600000977516174},{"id":"https://openalex.org/keywords/estimation","display_name":"Estimation","score":0.3328999876976013},{"id":"https://openalex.org/keywords/data-integration","display_name":"Data integration","score":0.33219999074935913},{"id":"https://openalex.org/keywords/maximum-likelihood","display_name":"Maximum likelihood","score":0.2888999879360199},{"id":"https://openalex.org/keywords/term","display_name":"Term (time)","score":0.2786000072956085}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.49140000343322754},{"id":"https://openalex.org/C207390915","wikidata":"https://www.wikidata.org/wiki/Q1230525","display_name":"Divergence (linguistics)","level":2,"score":0.45190000534057617},{"id":"https://openalex.org/C171752962","wikidata":"https://www.wikidata.org/wiki/Q255166","display_name":"Kullback\u2013Leibler divergence","level":2,"score":0.44600000977516174},{"id":"https://openalex.org/C33923547","wikidata":"https://www.wikidata.org/wiki/Q395","display_name":"Mathematics","level":0,"score":0.42149999737739563},{"id":"https://openalex.org/C11413529","wikidata":"https://www.wikidata.org/wiki/Q8366","display_name":"Algorithm","level":1,"score":0.3856000006198883},{"id":"https://openalex.org/C96250715","wikidata":"https://www.wikidata.org/wiki/Q965330","display_name":"Estimation","level":2,"score":0.3328999876976013},{"id":"https://openalex.org/C72634772","wikidata":"https://www.wikidata.org/wiki/Q386824","display_name":"Data integration","level":2,"score":0.33219999074935913},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.30329999327659607},{"id":"https://openalex.org/C49781872","wikidata":"https://www.wikidata.org/wiki/Q1045555","display_name":"Maximum likelihood","level":2,"score":0.2888999879360199},{"id":"https://openalex.org/C61797465","wikidata":"https://www.wikidata.org/wiki/Q1188986","display_name":"Term (time)","level":2,"score":0.2786000072956085},{"id":"https://openalex.org/C167928553","wikidata":"https://www.wikidata.org/wiki/Q1376021","display_name":"Estimation theory","level":2,"score":0.2773999869823456},{"id":"https://openalex.org/C107673813","wikidata":"https://www.wikidata.org/wiki/Q812534","display_name":"Bayesian probability","level":2,"score":0.27649998664855957},{"id":"https://openalex.org/C28826006","wikidata":"https://www.wikidata.org/wiki/Q33521","display_name":"Applied mathematics","level":1,"score":0.2655999958515167},{"id":"https://openalex.org/C124101348","wikidata":"https://www.wikidata.org/wiki/Q172491","display_name":"Data mining","level":1,"score":0.25839999318122864},{"id":"https://openalex.org/C2777212361","wikidata":"https://www.wikidata.org/wiki/Q5127848","display_name":"Class (philosophy)","level":2,"score":0.25600001215934753},{"id":"https://openalex.org/C2780009758","wikidata":"https://www.wikidata.org/wiki/Q6804172","display_name":"Measure (data warehouse)","level":2,"score":0.2558000087738037}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.18653/v1/2026.acl-long.405","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.acl-long.405","pdf_url":"https://aclanthology.org/2026.acl-long.405.pdf","source":null,"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"}],"best_oa_location":{"id":"doi:10.18653/v1/2026.acl-long.405","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.acl-long.405","pdf_url":"https://aclanthology.org/2026.acl-long.405.pdf","source":null,"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":[],"awards":[{"id":"https://openalex.org/G8884928098","display_name":null,"funder_award_id":"U24B20181","funder_id":"https://openalex.org/F4320321001","funder_display_name":"National Natural Science Foundation of China"}],"funders":[{"id":"https://openalex.org/F4320321001","display_name":"National Natural Science Foundation of China","ror":"https://ror.org/01h0zpd94"}],"has_content":{"pdf":true,"grobid_xml":true},"content_urls":{"pdf":"https://content.openalex.org/works/W7166841842.pdf","grobid_xml":"https://content.openalex.org/works/W7166841842.grobid-xml"},"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Kullback-Leibler":[0],"(KL)":[1],"divergence":[2,44],"regularization":[3,168],"is":[4],"essential":[5],"for":[6,121,169],"stabilizing":[7],"reinforcement":[8],"learning":[9],"from":[10,50,55],"human":[11],"feedback":[12],"(RLHF)":[13],"in":[14,104],"large":[15],"language":[16,79],"models":[17,130],"(LLMs),":[18],"yet":[19],"its":[20],"exact":[21,154],"computation":[22],"requires":[23],"summing":[24],"over":[25,85],"vocabularies":[26],"of":[27,78,93,108],"all":[28],"tokens,":[29,89],"incurring":[30],"prohibitive":[31],"memory":[32,100],"costs":[33],"during":[34],"training.Existing":[35],"stochastic":[36,139],"estimators":[37],"circumvent":[38],"this":[39],"bottleneck":[40],"by":[41,82],"estimating":[42],"KL":[43,71,167],"using":[45],"only":[46,86],"the":[47,51,75,87,94,153],"sampled":[48],"token":[49],"trajectory,":[52],"but":[53],"suffer":[54],"high":[56],"variance":[57],"(k":[58,64],"1":[59],")":[60],"or":[61],"systematic":[62],"bias":[63],"2":[65],").We":[66],"propose":[67],"TIKE":[68,90,136,150],"(Top-k":[69],"Importance-weighted":[70],"Estimator),":[72],"which":[73],"exploits":[74],"Zipfian":[76],"structure":[77],"model":[80],"distributions:":[81],"deterministically":[83],"integrating":[84],"top-k":[88],"captures":[91],"most":[92],"probability":[95],"mass":[96],"while":[97,141],"effectively":[98],"reducing":[99],"cost.To":[101],"ensure":[102],"correctness":[103],"off-policy":[105],"settings":[106],"characteristic":[107],"Group":[109],"Relative":[110],"Policy":[111],"Optimization":[112],"(GRPO),":[113],"we":[114],"incorporate":[115],"importance":[116],"sampling":[117],"weights":[118],"that":[119,135,149],"correct":[120],"distribution":[122],"shift":[123],"between":[124],"rollout":[125],"and":[126],"optimization":[127],"policies.Experiments":[128],"on":[129],"across":[131],"diverse":[132],"benchmarks":[133],"demonstrate":[134],"consistently":[137],"outperforms":[138],"baselines,":[140],"exhibiting":[142],"substantially":[143],"lower":[144],"gradient":[145],"variance.Our":[146],"analysis":[147],"reveals":[148],"closely":[151],"tracks":[152],"Rao-Blackwellized":[155],"estimator":[156],"with":[157],"nearzero":[158],"variance,":[159],"offering":[160],"a":[161],"practical":[162],"path":[163],"toward":[164],"stable,":[165],"memory-efficient":[166],"reasoning-intensive":[170],"LLMs":[171],"training.":[172]},"counts_by_year":[],"updated_date":"2026-07-29T14:22:42.915294","created_date":"2026-07-02T00:00:00"}