{"id":"https://openalex.org/W7154239169","doi":"https://doi.org/10.48550/arxiv.2604.10079","title":"Why Supervised Fine-Tuning Fails to Learn: A Systematic Study of Incomplete Learning in Large Language Models","display_name":"Why Supervised Fine-Tuning Fails to Learn: A Systematic Study of Incomplete Learning in Large Language Models","publication_year":2026,"publication_date":"2026-04-11","ids":{"openalex":"https://openalex.org/W7154239169","doi":"https://doi.org/10.48550/arxiv.2604.10079"},"language":null,"primary_location":{"id":"doi:10.48550/arxiv.2604.10079","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2604.10079","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":null,"license_id":null,"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.2604.10079","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5133598525","display_name":"Chao Xue","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Xue, Chao","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5133584615","display_name":"Yao Wang","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Wang, Yao","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5067658211","display_name":"Mengqiao Liu","orcid":"https://orcid.org/0000-0001-8426-3091"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Liu, Mengqiao","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5133561897","display_name":"Di Liang","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Liang, Di","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5133587881","display_name":"Xingsheng Han","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Han, Xingsheng","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5133572360","display_name":"Peiyang Liu","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Liu, Peiyang","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5133569838","display_name":"Xianjie Wu","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Wu, Xianjie","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5133585249","display_name":"Chenyao Lu","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Lu, Chenyao","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5133561850","display_name":"Lei Jiang","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Jiang, Lei","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5133600184","display_name":"Yu Lu","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Lu, Yu","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5133562699","display_name":"Haibo Shi","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Shi, Haibo","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5133604915","display_name":"Shuang Liang","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Liang, Shuang","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5133622260","display_name":"Minlong Peng","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Peng, Minlong","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5133608705","display_name":"Flora D. Salim","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Salim, Flora D.","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/T10028","display_name":"Topic Modeling","score":0.2904999852180481,"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/T10028","display_name":"Topic Modeling","score":0.2904999852180481,"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":0.1282999962568283,"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/T12026","display_name":"Explainable Artificial Intelligence (XAI)","score":0.09520000219345093,"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/inference","display_name":"Inference","score":0.6288999915122986},{"id":"https://openalex.org/keywords/forgetting","display_name":"Forgetting","score":0.579200029373169},{"id":"https://openalex.org/keywords/aggregate","display_name":"Aggregate (composite)","score":0.5329999923706055},{"id":"https://openalex.org/keywords/supervised-learning","display_name":"Supervised learning","score":0.4952999949455261},{"id":"https://openalex.org/keywords/language-model","display_name":"Language model","score":0.42820000648498535},{"id":"https://openalex.org/keywords/semi-supervised-learning","display_name":"Semi-supervised learning","score":0.35929998755455017},{"id":"https://openalex.org/keywords/structured-prediction","display_name":"Structured prediction","score":0.3411000072956085}],"concepts":[{"id":"https://openalex.org/C2776214188","wikidata":"https://www.wikidata.org/wiki/Q408386","display_name":"Inference","level":2,"score":0.6288999915122986},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.6200000047683716},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.616100013256073},{"id":"https://openalex.org/C7149132","wikidata":"https://www.wikidata.org/wiki/Q1377840","display_name":"Forgetting","level":2,"score":0.579200029373169},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.567300021648407},{"id":"https://openalex.org/C4679612","wikidata":"https://www.wikidata.org/wiki/Q866298","display_name":"Aggregate (composite)","level":2,"score":0.5329999923706055},{"id":"https://openalex.org/C136389625","wikidata":"https://www.wikidata.org/wiki/Q334384","display_name":"Supervised learning","level":3,"score":0.4952999949455261},{"id":"https://openalex.org/C137293760","wikidata":"https://www.wikidata.org/wiki/Q3621696","display_name":"Language model","level":2,"score":0.42820000648498535},{"id":"https://openalex.org/C58973888","wikidata":"https://www.wikidata.org/wiki/Q1041418","display_name":"Semi-supervised learning","level":2,"score":0.35929998755455017},{"id":"https://openalex.org/C22367795","wikidata":"https://www.wikidata.org/wiki/Q7625208","display_name":"Structured prediction","level":2,"score":0.3411000072956085},{"id":"https://openalex.org/C9357733","wikidata":"https://www.wikidata.org/wiki/Q6878417","display_name":"Missing data","level":2,"score":0.33959999680519104},{"id":"https://openalex.org/C158600405","wikidata":"https://www.wikidata.org/wiki/Q5054566","display_name":"Causal inference","level":2,"score":0.33410000801086426},{"id":"https://openalex.org/C51632099","wikidata":"https://www.wikidata.org/wiki/Q3985153","display_name":"Training set","level":2,"score":0.3330000042915344},{"id":"https://openalex.org/C195324797","wikidata":"https://www.wikidata.org/wiki/Q33742","display_name":"Natural language","level":2,"score":0.30709999799728394},{"id":"https://openalex.org/C2983448237","wikidata":"https://www.wikidata.org/wiki/Q1078276","display_name":"Language understanding","level":2,"score":0.2978000044822693},{"id":"https://openalex.org/C204321447","wikidata":"https://www.wikidata.org/wiki/Q30642","display_name":"Natural language processing","level":1,"score":0.26829999685287476},{"id":"https://openalex.org/C11671645","wikidata":"https://www.wikidata.org/wiki/Q5054567","display_name":"Causal model","level":2,"score":0.26350000500679016},{"id":"https://openalex.org/C2780801425","wikidata":"https://www.wikidata.org/wiki/Q5164392","display_name":"Construct (python library)","level":2,"score":0.2563000023365021}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.48550/arxiv.2604.10079","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2604.10079","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":null,"license_id":null,"version":null,"is_accepted":false,"is_published":null,"raw_source_name":null,"raw_type":"Preprint"}],"best_oa_location":{"id":"doi:10.48550/arxiv.2604.10079","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2604.10079","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":null,"license_id":null,"version":null,"is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"Preprint"},"sustainable_development_goals":[{"score":0.7267062067985535,"id":"https://metadata.un.org/sdg/4","display_name":"Quality Education"}],"awards":[],"funders":[],"has_content":{"grobid_xml":false,"pdf":false},"content_urls":null,"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Supervised":[0],"Fine-Tuning":[1],"(SFT)":[2],"is":[3,168],"the":[4,46,53,99,186],"standard":[5],"approach":[6],"for":[7,126,188],"adapting":[8],"large":[9],"language":[10],"models":[11,26],"(LLMs)":[12],"to":[13,29,42,68,140,196],"downstream":[14],"tasks.":[15],"However,":[16],"we":[17,86],"observe":[18],"a":[19,32,133],"persistent":[20,180],"failure":[21,67],"mode:":[22],"even":[23],"after":[24],"convergence,":[25],"often":[27],"fail":[28],"correctly":[30],"reproduce":[31],"subset":[33],"of":[34,57,91,191],"their":[35],"own":[36],"supervised":[37,70,193],"training":[38,145],"data.":[39],"We":[40,62,131],"refer":[41],"this":[43],"behavior":[44],"as":[45,65,155],"Incomplete":[47],"Learning":[48],"Phenomenon(ILP).":[49],"This":[50],"paper":[51],"presents":[52],"first":[54],"systematic":[55],"study":[56,150],"ILP":[58,64],"in":[59,98,175],"LLM":[60],"fine-tuning.":[61],"formalize":[63],"post-training":[66],"internalize":[69],"instances":[71],"and":[72,81,107,122,146,149,162,170,172,198],"demonstrate":[73],"its":[74],"prevalence":[75],"across":[76],"multiple":[77],"model":[78],"families,":[79],"domains,":[80],"datasets.":[82],"Through":[83],"controlled":[84],"analyses,":[85],"identify":[87],"five":[88],"recurrent":[89],"sources":[90],"incomplete":[92,166],"learning:":[93],"(1)":[94],"missing":[95],"prerequisite":[96],"knowledge":[97],"pre-trained":[100],"model,":[101],"(2)":[102],"conflicts":[103],"between":[104],"SFT":[105,114],"supervision":[106],"pre-training":[108],"knowledge,":[109],"(3)":[110],"internal":[111],"inconsistencies":[112],"within":[113],"data,":[115],"(4)":[116],"left-side":[117],"forgetting":[118],"during":[119],"sequential":[120],"fine-tuning,":[121],"(5)":[123],"insufficient":[124],"optimization":[125],"rare":[127],"or":[128],"complex":[129],"patterns.":[130],"introduce":[132],"diagnostic-first":[134],"framework":[135],"that":[136,165,173],"maps":[137],"unlearned":[138,181],"samples":[139],"these":[141],"causes":[142],"using":[143],"observable":[144],"inference":[147],"signals,":[148],"several":[151],"targeted":[152],"mitigation":[153],"strategies":[154],"causal":[156],"interventions.":[157],"Experiments":[158],"on":[159],"Qwen,":[160],"LLaMA,":[161],"OLMo2":[163],"show":[164],"learning":[167],"widespread":[169],"heterogeneous,":[171],"improvements":[174],"aggregate":[176],"metrics":[177],"can":[178],"mask":[179],"subsets.":[182],"The":[183],"findings":[184],"highlight":[185],"need":[187],"fine-grained":[189],"diagnosis":[190],"what":[192],"fine-tuning":[194],"fails":[195],"learn,":[197],"why.":[199]},"counts_by_year":[],"updated_date":"2026-07-01T06:00:48.157686","created_date":"2026-04-15T00:00:00"}