{"id":"https://openalex.org/W7154478539","doi":"https://doi.org/10.48550/arxiv.2604.11912","title":"How Transformers Learn to Plan via Multi-Token Prediction","display_name":"How Transformers Learn to Plan via Multi-Token Prediction","publication_year":2026,"publication_date":"2026-04-13","ids":{"openalex":"https://openalex.org/W7154478539","doi":"https://doi.org/10.48550/arxiv.2604.11912"},"language":null,"primary_location":{"id":"doi:10.48550/arxiv.2604.11912","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2604.11912","pdf_url":null,"source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":null,"issn":null,"is_oa":true,"is_in_doaj":false,"is_core":false,"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"},"type":"preprint","indexed_in":["datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://doi.org/10.48550/arxiv.2604.11912","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5124955561","display_name":"Jianhao Huang","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Huang, Jianhao","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5133708301","display_name":"Zhanpeng Zhou","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Zhou, Zhanpeng","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5133704361","display_name":"Renqiu Xia","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Xia, Renqiu","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5009166220","display_name":"Baharan Mirzasoleiman","orcid":"https://orcid.org/0000-0002-9374-9496"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Mirzasoleiman, Baharan","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5133652143","display_name":"Weijie J. Su","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Su, Weijie","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5133680682","display_name":"Wei Huang","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Huang, Wei","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":null,"has_fulltext":false,"cited_by_count":0,"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/T11714","display_name":"Multimodal Machine Learning Applications","score":0.4106999933719635,"subfield":{"id":"https://openalex.org/subfields/1707","display_name":"Computer Vision and Pattern Recognition"},"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/T11714","display_name":"Multimodal Machine Learning Applications","score":0.4106999933719635,"subfield":{"id":"https://openalex.org/subfields/1707","display_name":"Computer Vision and Pattern Recognition"},"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":0.17509999871253967,"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/T13702","display_name":"Machine Learning in Healthcare","score":0.06589999794960022,"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/transformer","display_name":"Transformer","score":0.5289000272750854},{"id":"https://openalex.org/keywords/boolean-satisfiability-problem","display_name":"Boolean satisfiability problem","score":0.4699000120162964},{"id":"https://openalex.org/keywords/graph","display_name":"Graph","score":0.42890000343322754},{"id":"https://openalex.org/keywords/satisfiability","display_name":"Satisfiability","score":0.3896999955177307},{"id":"https://openalex.org/keywords/causal-reasoning","display_name":"Causal reasoning","score":0.37700000405311584},{"id":"https://openalex.org/keywords/qualitative-reasoning","display_name":"Qualitative reasoning","score":0.35040000081062317},{"id":"https://openalex.org/keywords/case-based-reasoning","display_name":"Case-based reasoning","score":0.34630000591278076},{"id":"https://openalex.org/keywords/boosting","display_name":"Boosting (machine learning)","score":0.33000001311302185}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.6366999745368958},{"id":"https://openalex.org/C66322947","wikidata":"https://www.wikidata.org/wiki/Q11658","display_name":"Transformer","level":3,"score":0.5289000272750854},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.49869999289512634},{"id":"https://openalex.org/C6943359","wikidata":"https://www.wikidata.org/wiki/Q875276","display_name":"Boolean satisfiability problem","level":2,"score":0.4699000120162964},{"id":"https://openalex.org/C132525143","wikidata":"https://www.wikidata.org/wiki/Q141488","display_name":"Graph","level":2,"score":0.42890000343322754},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.3939000070095062},{"id":"https://openalex.org/C168773769","wikidata":"https://www.wikidata.org/wiki/Q1350299","display_name":"Satisfiability","level":2,"score":0.3896999955177307},{"id":"https://openalex.org/C115086926","wikidata":"https://www.wikidata.org/wiki/Q17004651","display_name":"Causal reasoning","level":3,"score":0.37700000405311584},{"id":"https://openalex.org/C83725634","wikidata":"https://www.wikidata.org/wiki/Q7268699","display_name":"Qualitative reasoning","level":2,"score":0.35040000081062317},{"id":"https://openalex.org/C20162079","wikidata":"https://www.wikidata.org/wiki/Q1151406","display_name":"Case-based reasoning","level":2,"score":0.34630000591278076},{"id":"https://openalex.org/C46686674","wikidata":"https://www.wikidata.org/wiki/Q466303","display_name":"Boosting (machine learning)","level":2,"score":0.33000001311302185},{"id":"https://openalex.org/C171268870","wikidata":"https://www.wikidata.org/wiki/Q1486676","display_name":"GRASP","level":2,"score":0.3285999894142151},{"id":"https://openalex.org/C80444323","wikidata":"https://www.wikidata.org/wiki/Q2878974","display_name":"Theoretical computer science","level":1,"score":0.32519999146461487},{"id":"https://openalex.org/C138673069","wikidata":"https://www.wikidata.org/wiki/Q322229","display_name":"Tracing","level":2,"score":0.3077000081539154},{"id":"https://openalex.org/C192209626","wikidata":"https://www.wikidata.org/wiki/Q190909","display_name":"Focus (optics)","level":2,"score":0.30079999566078186},{"id":"https://openalex.org/C41608201","wikidata":"https://www.wikidata.org/wiki/Q980509","display_name":"Embedding","level":2,"score":0.2904999852180481},{"id":"https://openalex.org/C164155591","wikidata":"https://www.wikidata.org/wiki/Q2067766","display_name":"Satisfiability modulo theories","level":2,"score":0.28450000286102295},{"id":"https://openalex.org/C136886441","wikidata":"https://www.wikidata.org/wiki/Q926129","display_name":"Normalization (sociology)","level":2,"score":0.2802000045776367},{"id":"https://openalex.org/C2777735758","wikidata":"https://www.wikidata.org/wiki/Q817765","display_name":"Path (computing)","level":2,"score":0.27070000767707825},{"id":"https://openalex.org/C177212765","wikidata":"https://www.wikidata.org/wiki/Q627335","display_name":"Workflow","level":2,"score":0.2648000121116638},{"id":"https://openalex.org/C189950617","wikidata":"https://www.wikidata.org/wiki/Q937228","display_name":"Property (philosophy)","level":2,"score":0.25619998574256897},{"id":"https://openalex.org/C121704057","wikidata":"https://www.wikidata.org/wiki/Q352070","display_name":"Collision","level":2,"score":0.25290000438690186},{"id":"https://openalex.org/C51632099","wikidata":"https://www.wikidata.org/wiki/Q3985153","display_name":"Training set","level":2,"score":0.2524000108242035}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.48550/arxiv.2604.11912","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2604.11912","pdf_url":null,"source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":null,"issn":null,"is_oa":true,"is_in_doaj":false,"is_core":false,"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":"doi:10.48550/arxiv.2604.11912","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2604.11912","pdf_url":null,"source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":null,"issn":null,"is_oa":true,"is_in_doaj":false,"is_core":false,"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"},"sustainable_development_goals":[],"awards":[],"funders":[],"has_content":{"pdf":false,"grobid_xml":false},"content_urls":null,"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"While":[0],"next-token":[1],"prediction":[2,24],"(NTP)":[3],"has":[4,26],"been":[5],"the":[6,102,107,113],"standard":[7],"objective":[8],"for":[9],"training":[10,134],"language":[11],"models,":[12],"it":[13],"often":[14],"struggles":[15],"to":[16,106,137],"capture":[17],"global":[18],"structure":[19],"in":[20],"reasoning":[21,71,100,153],"tasks.":[22],"Multi-token":[23],"(MTP)":[25],"recently":[27],"emerged":[28],"as":[29,74],"a":[30,50,83,88,97,124,132],"promising":[31],"alternative,":[32],"yet":[33],"its":[34],"underlying":[35],"mechanisms":[36],"remain":[37],"poorly":[38],"understood.":[39],"In":[40],"this":[41],"paper,":[42],"we":[43,55,81],"study":[44],"how":[45,143],"MTP":[46,58,95],"facilitates":[47],"reasoning,":[48],"with":[49],"focus":[51],"on":[52,62,87],"planning.":[53],"Empirically,":[54],"show":[56],"that":[57,94],"consistently":[59],"outperforms":[60],"NTP":[61],"both":[63],"synthetic":[64],"graph":[65,90],"path-finding":[66],"tasks":[67],"and":[68,76,110,151],"more":[69],"realistic":[70],"benchmarks,":[72],"such":[73],"Countdown":[75],"boolean":[77],"satisfiability":[78],"problems.":[79],"Theoretically,":[80],"analyze":[82],"simplified":[84],"two-layer":[85],"Transformer":[86],"star":[89],"task.":[91],"We":[92],"prove":[93],"induces":[96],"two-stage":[98],"reverse":[99],"process:":[101],"model":[103],"first":[104],"attends":[105],"end":[108],"node":[109],"then":[111],"reconstructs":[112],"path":[114],"by":[115],"tracing":[116],"intermediate":[117],"nodes":[118],"backward.":[119],"This":[120],"behavior":[121],"arises":[122],"from":[123],"gradient":[125],"decoupling":[126],"property":[127],"of":[128],"MTP,":[129],"which":[130],"provides":[131],"cleaner":[133],"signal":[135],"compared":[136],"NTP.":[138],"Ultimately,":[139],"our":[140],"results":[141],"highlight":[142],"multi-token":[144],"objectives":[145],"inherently":[146],"bias":[147],"optimization":[148],"toward":[149],"robust":[150],"interpretable":[152],"circuits.":[154]},"counts_by_year":[],"updated_date":"2026-07-01T06:00:48.157686","created_date":"2026-04-16T00:00:00"}