{"id":"https://openalex.org/W4401103013","doi":"https://doi.org/10.48550/arxiv.2402.19446","title":"ArCHer: Training Language Model Agents via Hierarchical Multi-Turn RL","display_name":"ArCHer: Training Language Model Agents via Hierarchical Multi-Turn RL","publication_year":2024,"publication_date":"2024-02-29","ids":{"openalex":"https://openalex.org/W4401103013","doi":"https://doi.org/10.48550/arxiv.2402.19446"},"language":"en","primary_location":{"id":"pmh:oai:arXiv.org:2402.19446","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2402.19446","pdf_url":"https://arxiv.org/pdf/2402.19446","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/2402.19446","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5100645656","display_name":"Yifei Zhou","orcid":"https://orcid.org/0000-0002-2176-9316","observed_orcids":["https://orcid.org/0000-0002-2176-9316","https://orcid.org/0000-0002-6372-8205","https://orcid.org/0009-0004-6960-8281"]},"institutions":[{"id":"https://openalex.org/I95457486","display_name":"University of California, Berkeley","ror":"https://ror.org/01an7q238","country_code":"US","type":"education","lineage":["https://openalex.org/I95457486"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Zhou, Yifei","raw_affiliation_strings":["University of California , Berkeley ,"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"University of California , Berkeley ,","institution_ids":["https://openalex.org/I95457486"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5055456006","display_name":"Andrea Zanette","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I95457486","display_name":"University of California, Berkeley","ror":"https://ror.org/01an7q238","country_code":"US","type":"education","lineage":["https://openalex.org/I95457486"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Zanette, Andrea","raw_affiliation_strings":["University of California , Berkeley ,"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"University of California , Berkeley ,","institution_ids":["https://openalex.org/I95457486"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5112313376","display_name":"Jiayi Pan","orcid":"https://orcid.org/0009-0002-3432-855X","observed_orcids":["https://orcid.org/0009-0002-3432-855X"]},"institutions":[{"id":"https://openalex.org/I95457486","display_name":"University of California, Berkeley","ror":"https://ror.org/01an7q238","country_code":"US","type":"education","lineage":["https://openalex.org/I95457486"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Pan, Jiayi","raw_affiliation_strings":["University of California , Berkeley ,"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"University of California , Berkeley ,","institution_ids":["https://openalex.org/I95457486"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5026322200","display_name":"Sergey Levine","orcid":"https://orcid.org/0000-0001-6764-2743","observed_orcids":["https://orcid.org/0000-0001-6764-2743"]},"institutions":[{"id":"https://openalex.org/I95457486","display_name":"University of California, Berkeley","ror":"https://ror.org/01an7q238","country_code":"US","type":"education","lineage":["https://openalex.org/I95457486"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Levine, Sergey","raw_affiliation_strings":["University of California , Berkeley ,"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"University of California , Berkeley ,","institution_ids":["https://openalex.org/I95457486"]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5102493293","display_name":"Aviral Kumar","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I4210090411","display_name":"Google DeepMind (United Kingdom)","ror":"https://ror.org/00971b260","country_code":"GB","type":"company","lineage":["https://openalex.org/I4210090411","https://openalex.org/I4210128969"]}],"countries":["GB"],"is_corresponding":false,"raw_author_name":"Kumar, Aviral","raw_affiliation_strings":["Google DeepMind"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Google DeepMind","institution_ids":["https://openalex.org/I4210090411"]}]}],"institutions":[],"countries_distinct_count":2,"institutions_distinct_count":2,"corresponding_author_ids":[],"corresponding_institution_ids":[],"apc_list":null,"apc_paid":null,"fwci":null,"has_fulltext":true,"cited_by_count":2,"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/T10462","display_name":"Reinforcement Learning in Robotics","score":0.9423999786376953,"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/T10462","display_name":"Reinforcement Learning in Robotics","score":0.9423999786376953,"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":0.010300000198185444,"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":0.010300000198185444,"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/llm-based-agents","display_name":"LLM agents","score":0.7139999866485596},{"id":"https://openalex.org/keywords/multi-turn-reinforcement-learning","display_name":"multi-turn reinforcement learning","score":0.9580000042915344},{"id":"https://openalex.org/keywords/hierarchical-reinforcement-learning","display_name":"hierarchical reinforcement learning","score":0.9940000176429749},{"id":"https://openalex.org/keywords/llm-fine-tuning","display_name":"LLM fine-tuning","score":0.7979999780654907},{"id":"https://openalex.org/keywords/goal-directed-decision-making","display_name":"goal-directed decision making","score":0.6729999780654907},{"id":"https://openalex.org/keywords/delayed-rewards","display_name":"delayed rewards","score":0.6359999775886536},{"id":"https://openalex.org/keywords/credit-assignment","display_name":"credit assignment","score":0.671999990940094},{"id":"https://openalex.org/keywords/sample-efficiency","display_name":"sample efficiency","score":0.8090000152587891},{"id":"https://openalex.org/keywords/archers","display_name":"archers","score":0.7820000052452087}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.6039972901344299},{"id":"https://openalex.org/C85641259","wikidata":"https://www.wikidata.org/wiki/Q290042","display_name":"Turn (biochemistry)","level":2,"score":0.5393374562263489},{"id":"https://openalex.org/C2777211547","wikidata":"https://www.wikidata.org/wiki/Q17141490","display_name":"Training (meteorology)","level":2,"score":0.5213112235069275},{"id":"https://openalex.org/C137293760","wikidata":"https://www.wikidata.org/wiki/Q3621696","display_name":"Language model","level":2,"score":0.44907456636428833},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.3264278769493103},{"id":"https://openalex.org/C185592680","wikidata":"https://www.wikidata.org/wiki/Q2329","display_name":"Chemistry","level":0,"score":0.05631348490715027},{"id":"https://openalex.org/C205649164","wikidata":"https://www.wikidata.org/wiki/Q1071","display_name":"Geography","level":0,"score":0.043422430753707886},{"id":"https://openalex.org/C55493867","wikidata":"https://www.wikidata.org/wiki/Q7094","display_name":"Biochemistry","level":1,"score":0.0},{"id":"https://openalex.org/C153294291","wikidata":"https://www.wikidata.org/wiki/Q25261","display_name":"Meteorology","level":1,"score":0.0}],"mesh":[],"locations_count":2,"locations":[{"id":"pmh:oai:arXiv.org:2402.19446","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2402.19446","pdf_url":"https://arxiv.org/pdf/2402.19446","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.2402.19446","doi":"https://doi.org/10.48550/arxiv.2402.19446","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2402.19446","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:2402.19446","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2402.19446","pdf_url":"https://arxiv.org/pdf/2402.19446","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":[],"study_designs":[],"awards":[{"id":"https://openalex.org/G1559816056","display_name":"RI: Small: Extracting Knowledge from Language Models for Decision Making","funder_award_id":"2246811","funder_id":"https://openalex.org/F4320306076","funder_display_name":"National Science Foundation"},{"id":"https://openalex.org/G8876996369","display_name":null,"funder_award_id":"N00014","funder_id":"https://openalex.org/F4320337345","funder_display_name":"Office of Naval Research"}],"funders":[{"id":"https://openalex.org/F4320306076","display_name":"National Science Foundation","ror":"https://ror.org/021nxhr62"},{"id":"https://openalex.org/F4320334164","display_name":"Simons Institute for the Theory of Computing, University of California Berkeley","ror":null},{"id":"https://openalex.org/F4320337345","display_name":"Office of Naval Research","ror":"https://ror.org/00rk2pe57"}],"has_content":{"pdf":true,"grobid_xml":true},"content_urls":{"pdf":"https://content.openalex.org/works/W4401103013.pdf","grobid_xml":"https://content.openalex.org/works/W4401103013.grobid-xml"},"referenced_works_count":0,"referenced_works":[],"related_works":["https://openalex.org/W4389260861","https://openalex.org/W4414971194","https://openalex.org/W7130605930","https://openalex.org/W4417167303","https://openalex.org/W4414810684","https://openalex.org/W7166839066","https://openalex.org/W6891877852","https://openalex.org/W7102410087","https://openalex.org/W7196976016","https://openalex.org/W7140105616"],"abstract_inverted_index":{"A":[0],"broad":[1],"use":[2],"case":[3],"of":[4,109,151,256],"large":[5],"language":[6],"models":[7],"(LLMs)":[8],"is":[9],"in":[10,113,187],"goal-directed":[11],"decision-making":[12],"tasks":[13],"(or":[14],"\"agent\"":[15],"tasks),":[16],"where":[17],"an":[18],"LLM":[19],"needs":[20],"to":[21,39,62,91,195,211,235],"not":[22],"just":[23],"generate":[24],"completions":[25],"for":[26,71,130,139,144,156],"a":[27,36,41,59,137,178,189,201,213,227,253],"given":[28],"prompt,":[29],"but":[30,67],"rather":[31],"make":[32],"intelligent":[33],"decisions":[34],"over":[35,95,198,259],"multi-turn":[37,127,141],"interaction":[38],"accomplish":[40],"task":[42],"(e.g.,":[43,158],"when":[44],"interacting":[45],"with":[46,88,226,265],"the":[47,89,118,149,270],"web,":[48],"using":[49],"tools,":[50],"or":[51,101,219],"providing":[52],"customer":[53],"support).":[54],"Reinforcement":[55],"learning":[56],"(RL)":[57],"provides":[58],"general":[60],"paradigm":[61],"address":[63],"such":[64],"agent":[65,114,250],"tasks,":[66,251],"current":[68],"RL":[69,83,128,142,154,180,185,193,203,237],"methods":[70,84,155],"LLMs":[72,87,157],"largely":[73],"focus":[74],"on":[75,249],"optimizing":[76],"single-turn":[77,82,153],"rewards.":[78],"By":[79],"construction,":[80],"most":[81],"cannot":[85],"endow":[86],"ability":[90],"intelligently":[92],"seek":[93],"information":[94],"multiple":[96,164],"turns,":[97,165],"perform":[98],"credit":[99],"assignment,":[100],"reason":[102],"about":[103,257],"their":[104],"past":[105],"actions":[106],"--":[107],"all":[108],"which":[110],"are":[111],"critical":[112],"tasks.":[115],"This":[116],"raises":[117],"question:":[119],"how":[120],"can":[121,231],"we":[122,135,240,275],"design":[123],"effective":[124],"and":[125,168,182,200,247],"efficient":[126],"algorithms":[129,143,186],"LLMs?":[131],"In":[132],"this":[133,207],"paper,":[134],"develop":[136],"framework":[138,176],"building":[140],"fine-tuning":[145],"LLMs,":[146],"that":[147,205,242,274],"preserves":[148],"flexibility":[150],"existing":[152,260],"proximal":[159],"policy":[160,215],"optimization),":[161],"while":[162,262],"accommodating":[163],"long":[166],"horizons,":[167],"delayed":[169],"rewards":[170],"effectively.":[171],"To":[172],"do":[173],"this,":[174],"our":[175],"adopts":[177],"hierarchical":[179,222],"approach":[181],"runs":[183],"two":[184],"parallel:":[188],"high-level":[190,208],"off-policy":[191],"value-based":[192],"algorithm":[194,204],"aggregate":[196],"reward":[197],"utterances,":[199],"low-level":[202],"utilizes":[206],"value":[209],"function":[210],"train":[212],"token":[214],"within":[216],"each":[217],"utterance":[218],"turn.":[220],"Our":[221],"framework,":[223],"Actor-Critic":[224],"Framework":[225],"Hierarchical":[228],"Structure":[229],"(ArCHer),":[230],"also":[232,263],"give":[233],"rise":[234],"other":[236],"methods.":[238],"Empirically,":[239],"find":[241],"ArCHer":[243],"significantly":[244],"improves":[245],"efficiency":[246,255],"performance":[248],"attaining":[252],"sample":[254],"100x":[258],"methods,":[261],"improving":[264],"larger":[266],"model":[267],"capacity":[268],"(upto":[269],"7":[271],"billion":[272],"scale":[273],"tested":[276],"on).":[277]},"counts_by_year":[{"year":2025,"cited_by_count":2}],"updated_date":"2026-09-30T09:27:27.532063","created_date":"2024-07-31T00:00:00"}