{"id":"https://openalex.org/W7204883073","doi":"https://doi.org/10.18653/2026.findings-emnlp.1651","title":"How do World Models and Policies Compose in LLM Agents? A Joint Spectral and Behavioral Account","display_name":"How do World Models and Policies Compose in LLM Agents? A Joint Spectral and Behavioral Account","publication_year":2026,"publication_date":"2026-01-01","ids":{"openalex":"https://openalex.org/W7204883073","doi":"https://doi.org/10.18653/2026.findings-emnlp.1651"},"language":"en","primary_location":{"id":"doi:10.18653/2026.findings-emnlp.1651","doi":"https://doi.org/10.18653/2026.findings-emnlp.1651","is_oa":true,"landing_page_url":"https://doi.org/10.18653/2026.findings-emnlp.1651","pdf_url":"https://aclanthology.org/2026.findings-emnlp.1651.pdf","source":{"id":"https://openalex.org/S7407096085","display_name":"Findings of the Association for Computational Linguistics: EMNLP (Findings of EMNLP)","issn_l":null,"issn":null,"is_oa":false,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":"https://openalex.org/P4310320244","host_organization_name":"Association for Computational Linguistics","host_organization_lineage":["https://openalex.org/P4310320244"],"host_organization_lineage_names":["Association for Computational Linguistics"],"type":"conference"},"license":null,"license_id":null,"version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Findings of the Association for Computational Linguistics: EMNLP 2026","raw_type":"proceedings-article"},"volume":null,"type":"conference-paper","indexed_in":["arxiv","crossref","datacite"],"open_access":{"is_oa":true,"oa_status":"bronze","oa_url":"https://aclanthology.org/2026.findings-emnlp.1651.pdf","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5135915403","display_name":"Ruize Xu","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I107672454","display_name":"Dartmouth College","ror":"https://ror.org/049s0rh22","country_code":"US","type":"education","lineage":["https://openalex.org/I107672454"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Ruize Xu","raw_affiliation_strings":["Dartmouth College"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Dartmouth College","institution_ids":["https://openalex.org/I107672454"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5148863736","display_name":"Xiao Yu","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I78577930","display_name":"Columbia University","ror":"https://ror.org/00hj8s172","country_code":"US","type":"education","lineage":["https://openalex.org/I78577930"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Xiao Yu","raw_affiliation_strings":["Columbia University"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Columbia University","institution_ids":["https://openalex.org/I78577930"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5148938811","display_name":"Yujin Tang","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I107672454","display_name":"Dartmouth College","ror":"https://ror.org/049s0rh22","country_code":"US","type":"education","lineage":["https://openalex.org/I107672454"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Yujin Tang","raw_affiliation_strings":["Dartmouth College"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Dartmouth College","institution_ids":["https://openalex.org/I107672454"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5023899239","display_name":"Chenming Shang","orcid":"https://orcid.org/0009-0004-8945-2165","observed_orcids":["https://orcid.org/0009-0004-8945-2165"]},"institutions":[{"id":"https://openalex.org/I107672454","display_name":"Dartmouth College","ror":"https://ror.org/049s0rh22","country_code":"US","type":"education","lineage":["https://openalex.org/I107672454"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Chenming Shang","raw_affiliation_strings":["Dartmouth College"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Dartmouth College","institution_ids":["https://openalex.org/I107672454"]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5148868405","display_name":"Nikhil Singh","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I107672454","display_name":"Dartmouth College","ror":"https://ror.org/049s0rh22","country_code":"US","type":"education","lineage":["https://openalex.org/I107672454"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Nikhil Singh","raw_affiliation_strings":["Dartmouth College"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Dartmouth College","institution_ids":["https://openalex.org/I107672454"]}]}],"institutions":[],"countries_distinct_count":1,"institutions_distinct_count":2,"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":0.69533853,"is_in_top_1_percent":false,"is_in_top_10_percent":false},"cited_by_percentile_year":null,"biblio":{"volume":null,"issue":null,"first_page":"35162","last_page":"35182"},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T10462","display_name":"Reinforcement Learning in Robotics","score":0.989300012588501,"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":0.989300012588501,"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/T11307","display_name":"Domain Adaptation and Few-Shot Learning","score":0.0017000000225380063,"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/T10906","display_name":"AI-based Problem Solving and Planning","score":0.00139999995008111,"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":0.8130000233650208},{"id":"https://openalex.org/keywords/world-models","display_name":"world models","score":0.9860000014305115},{"id":"https://openalex.org/keywords/next-state-prediction","display_name":"next-state prediction","score":0.8399999737739563},{"id":"https://openalex.org/keywords/policy-optimization","display_name":"policy optimization","score":0.5820000171661377},{"id":"https://openalex.org/keywords/reward-maximization","display_name":"reward maximization","score":0.6589999794960022},{"id":"https://openalex.org/keywords/spectral-geometry","display_name":"spectral geometry","score":0.5270000100135803},{"id":"https://openalex.org/keywords/exploration","display_name":"exploration","score":0.3499999940395355},{"id":"https://openalex.org/keywords/sequential-training","display_name":"sequential training","score":0.41600000858306885}],"concepts":[{"id":"https://openalex.org/C63479239","wikidata":"https://www.wikidata.org/wiki/Q7353546","display_name":"Robustness (evolution)","level":3,"score":0.6933000087738037},{"id":"https://openalex.org/C32834561","wikidata":"https://www.wikidata.org/wiki/Q660730","display_name":"Subspace topology","level":2,"score":0.6549999713897705},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.6344000101089478},{"id":"https://openalex.org/C177264268","wikidata":"https://www.wikidata.org/wiki/Q1514741","display_name":"Set (abstract data type)","level":2,"score":0.5562999844551086},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.4943000078201294},{"id":"https://openalex.org/C51632099","wikidata":"https://www.wikidata.org/wiki/Q3985153","display_name":"Training set","level":2,"score":0.4706000089645386},{"id":"https://openalex.org/C18555067","wikidata":"https://www.wikidata.org/wiki/Q8375051","display_name":"Joint (building)","level":2,"score":0.4553000032901764},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.42089998722076416},{"id":"https://openalex.org/C2779110517","wikidata":"https://www.wikidata.org/wiki/Q1240788","display_name":"Supervisor","level":2,"score":0.4052000045776367},{"id":"https://openalex.org/C2777211547","wikidata":"https://www.wikidata.org/wiki/Q17141490","display_name":"Training (meteorology)","level":2,"score":0.39890000224113464},{"id":"https://openalex.org/C2779436431","wikidata":"https://www.wikidata.org/wiki/Q30672407","display_name":"Policy learning","level":2,"score":0.3903999924659729},{"id":"https://openalex.org/C57493831","wikidata":"https://www.wikidata.org/wiki/Q3134666","display_name":"Projection (relational algebra)","level":2,"score":0.382099986076355},{"id":"https://openalex.org/C204323151","wikidata":"https://www.wikidata.org/wiki/Q905424","display_name":"Range (aeronautics)","level":2,"score":0.36059999465942383},{"id":"https://openalex.org/C165696696","wikidata":"https://www.wikidata.org/wiki/Q11287","display_name":"Exploit","level":2,"score":0.30660000443458557},{"id":"https://openalex.org/C97541855","wikidata":"https://www.wikidata.org/wiki/Q830687","display_name":"Reinforcement learning","level":2,"score":0.267300009727478},{"id":"https://openalex.org/C107457646","wikidata":"https://www.wikidata.org/wiki/Q207434","display_name":"Human\u2013computer interaction","level":1,"score":0.2583000063896179},{"id":"https://openalex.org/C113843644","wikidata":"https://www.wikidata.org/wiki/Q901882","display_name":"Interface (matter)","level":4,"score":0.2547999918460846}],"mesh":[],"locations_count":3,"locations":[{"id":"doi:10.18653/2026.findings-emnlp.1651","doi":"https://doi.org/10.18653/2026.findings-emnlp.1651","is_oa":true,"landing_page_url":"https://doi.org/10.18653/2026.findings-emnlp.1651","pdf_url":"https://aclanthology.org/2026.findings-emnlp.1651.pdf","source":{"id":"https://openalex.org/S7407096085","display_name":"Findings of the Association for Computational Linguistics: EMNLP (Findings of EMNLP)","issn_l":null,"issn":null,"is_oa":false,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":"https://openalex.org/P4310320244","host_organization_name":"Association for Computational Linguistics","host_organization_lineage":["https://openalex.org/P4310320244"],"host_organization_lineage_names":["Association for Computational Linguistics"],"type":"conference"},"license":null,"license_id":null,"version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Findings of the Association for Computational Linguistics: EMNLP 2026","raw_type":"proceedings-article"},{"id":"pmh:oai:arXiv.org:2608.30067","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2608.30067","pdf_url":"https://arxiv.org/pdf/2608.30067","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":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},{"id":"doi:10.48550/arxiv.2608.30067","doi":"https://doi.org/10.48550/arxiv.2608.30067","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2608.30067","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":"doi:10.18653/2026.findings-emnlp.1651","doi":"https://doi.org/10.18653/2026.findings-emnlp.1651","is_oa":true,"landing_page_url":"https://doi.org/10.18653/2026.findings-emnlp.1651","pdf_url":"https://aclanthology.org/2026.findings-emnlp.1651.pdf","source":{"id":"https://openalex.org/S7407096085","display_name":"Findings of the Association for Computational Linguistics: EMNLP (Findings of EMNLP)","issn_l":null,"issn":null,"is_oa":false,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":"https://openalex.org/P4310320244","host_organization_name":"Association for Computational Linguistics","host_organization_lineage":["https://openalex.org/P4310320244"],"host_organization_lineage_names":["Association for Computational Linguistics"],"type":"conference"},"license":null,"license_id":null,"version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Findings of the Association for Computational Linguistics: EMNLP 2026","raw_type":"proceedings-article"},"sustainable_development_goals":[],"study_designs":[],"awards":[],"funders":[],"has_content":{"pdf":true,"grobid_xml":true},"content_urls":{"pdf":"https://content.openalex.org/works/W7204883073.pdf","grobid_xml":"https://content.openalex.org/works/W7204883073.grobid-xml"},"referenced_works_count":0,"referenced_works":[],"related_works":["https://openalex.org/W7117153217","https://openalex.org/W7127888457","https://openalex.org/W7165941689","https://openalex.org/W7163159075","https://openalex.org/W7206159767","https://openalex.org/W4403364427","https://openalex.org/W4414633118","https://openalex.org/W7216147366","https://openalex.org/W4416034645","https://openalex.org/W7166590586"],"abstract_inverted_index":{"How":[0],"do":[1],"LLM":[2],"agents":[3],"come":[4],"to":[5,60,182],"both":[6,153],"understand":[7],"environments":[8],"they":[9],"act":[10],"in":[11,73,169],"and":[12,25,50,114,151,163,173],"master":[13],"tasks":[14],"set":[15],"within":[16],"them?Through":[17],"controlled":[18],"experiments":[19],"combining":[20],"world-model":[21,46,146],"training":[22,27,122],"(next-state":[23],"prediction)":[24],"policy":[26,56,84,121,149],"(reward":[28],"maximization),":[29],"we":[30,43,70,102,118],"investigate":[31],"this":[32,132],"question.We":[33],"dissect":[34],"the":[35,76,88,104,138,156,184],"resulting":[36],"models":[37],"through":[38],"their":[39],"additive":[40],"parameter":[41],"updates.Geometrically,":[42],"find":[44,71,103],"effective":[45],"updates":[47,57],"are":[48],"low-rank":[49],"share":[51],"an":[52,144],"input-feature":[53],"subspace":[54],"with":[55,133],"while":[58],"writing":[59],"nearly":[61],"orthogonal":[62],"output":[63],"directions,":[64,93],"whether":[65],"trained":[66,106],"separately":[67],"or":[68],"sequentially.However,":[69],"that,":[72],"projection":[74],"interventions,":[75],"sequential":[77],"update":[78],"induces":[79],"more":[80],"robustness":[81],"than":[82],"separate":[83],"RL":[85],"when":[86],"removing":[87],"world":[89,124,161],"model's":[90],"leading":[91],"input":[92,100,141],"suggesting":[94],"that":[95,174],"it":[96,129],"has":[97],"learned":[98,168],"alternative":[99],"pathways.Behaviorally,":[101],"sequentially":[105],"agent":[107],"explores":[108],"a":[109],"wider":[110],"range":[111],"of":[112],"states":[113],"actions.Based":[115],"on":[116,137],"this,":[117],"ask:":[119],"does":[120],"preserve":[123],"knowledge":[125,162],"as":[126,128],"well":[127],"could?We":[130],"probe":[131],"trainingfree":[134],"merging":[135],"built":[136],"geometrically":[139,170],"motivated":[140],"basis":[142],"plus":[143],"online":[145],"loss":[147],"during":[148],"RL,":[150],"show":[152],"improve":[154],"over":[155],"untreated":[157],"baseline.Our":[158],"findings":[159],"suggest":[160],"task-directed":[164],"ability":[165],"can":[166],"be":[167],"complementary":[171],"forms,":[172],"future":[175],"post-training":[176],"pipelines":[177],"should":[178],"consider":[179],"how":[180],"best":[181],"engineer":[183],"interface":[185],"between":[186],"them.":[187]},"counts_by_year":[],"updated_date":"2026-10-31T12:09:11.096373","created_date":"2026-09-02T00:00:00"}