{"id":"https://openalex.org/W6947817954","doi":"https://doi.org/10.48550/arxiv.2502.20681","title":"Disentangling Feature Structure: A Mathematically Provable Two-Stage Training Dynamics in Transformers","display_name":"Disentangling Feature Structure: A Mathematically Provable Two-Stage Training Dynamics in Transformers","publication_year":2025,"publication_date":"2025-02-28","ids":{"openalex":"https://openalex.org/W6947817954","doi":"https://doi.org/10.48550/arxiv.2502.20681"},"language":"en","primary_location":{"id":"pmh:oai:arXiv.org:2502.20681","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2502.20681","pdf_url":"https://arxiv.org/pdf/2502.20681","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"},"volume":null,"type":"preprint","indexed_in":["arxiv","datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://arxiv.org/pdf/2502.20681","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":null,"display_name":"Gong, Zixuan","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I62853816","display_name":"Beijing University of Civil Engineering and Architecture","ror":"https://ror.org/02yj0p855","country_code":"CN","type":"education","lineage":["https://openalex.org/I62853816"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Gong, Zixuan","raw_affiliation_strings":["School of Mechanical-Electronic and Vehicle Engineering , Beijing University of Civil Engineering and Architecture , Beijing , 102616 , China"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"School of Mechanical-Electronic and Vehicle Engineering , Beijing University of Civil Engineering and Architecture , Beijing , 102616 , China","institution_ids":["https://openalex.org/I62853816"]}]},{"author_position":"middle","author":{"id":null,"display_name":"Li, Shijia","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I181679659","display_name":"Shanghai University of Finance and Economics","ror":"https://ror.org/00wtvfq62","country_code":"CN","type":"education","lineage":["https://openalex.org/I181679659"]},{"id":"https://openalex.org/I78988378","display_name":"Renmin University of China","ror":"https://ror.org/041pakw92","country_code":"CN","type":"education","lineage":["https://openalex.org/I78988378"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Li, Shijia","raw_affiliation_strings":["Gaoling School of Artificial Intelligence , Renmin University of China , Beijing 100872 , China","School of Statistics and Management , Shanghai University of Finance and Economics , Shanghai , 200433 , China"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Gaoling School of Artificial Intelligence , Renmin University of China , Beijing 100872 , China","institution_ids":["https://openalex.org/I78988378"]},{"raw_affiliation_string":"School of Statistics and Management , Shanghai University of Finance and Economics , Shanghai , 200433 , China","institution_ids":["https://openalex.org/I181679659"]}]},{"author_position":"middle","author":{"id":null,"display_name":"Liu, Yong","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I78988378","display_name":"Renmin University of China","ror":"https://ror.org/041pakw92","country_code":"CN","type":"education","lineage":["https://openalex.org/I78988378"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Liu, Yong","raw_affiliation_strings":["Gaoling School of Artificial Intelligence , Renmin University of China , Beijing 100872 , China"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Gaoling School of Artificial Intelligence , Renmin University of China , Beijing 100872 , China","institution_ids":["https://openalex.org/I78988378"]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5015535373","display_name":"Jiaye Teng","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I78988378","display_name":"Renmin University of China","ror":"https://ror.org/041pakw92","country_code":"CN","type":"education","lineage":["https://openalex.org/I78988378"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Teng, Jiaye","raw_affiliation_strings":["Gaoling School of Artificial Intelligence , Renmin University of China , Beijing 100872 , China"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Gaoling School of Artificial Intelligence , Renmin University of China , Beijing 100872 , China","institution_ids":["https://openalex.org/I78988378"]}]}],"institutions":[],"countries_distinct_count":1,"institutions_distinct_count":3,"corresponding_author_ids":[],"corresponding_institution_ids":[],"apc_list":null,"apc_paid":null,"fwci":null,"has_fulltext":true,"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":true,"primary_topic":{"id":"https://openalex.org/T10028","display_name":"Topic Modeling","score":0.6596999764442444,"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":0.6596999764442444,"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/T11612","display_name":"Stochastic Gradient Optimization Techniques","score":0.20469999313354492,"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/T10320","display_name":"Neural Networks and Applications","score":0.04360000044107437,"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/feature-disentanglement","display_name":"feature disentanglement","score":0.7929999828338623},{"id":"https://openalex.org/keywords/transformer","display_name":"transformer models","score":0.6549999713897705},{"id":"https://openalex.org/keywords/gpt-2","display_name":"GPT-2","score":0.7860000133514404},{"id":"https://openalex.org/keywords/structured-data","display_name":"structured data","score":0.49900001287460327}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.6858999729156494},{"id":"https://openalex.org/C2776401178","wikidata":"https://www.wikidata.org/wiki/Q12050496","display_name":"Feature (linguistics)","level":2,"score":0.6327000260353088},{"id":"https://openalex.org/C66322947","wikidata":"https://www.wikidata.org/wiki/Q11658","display_name":"Transformer","level":3,"score":0.6086000204086304},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.6029000282287598},{"id":"https://openalex.org/C60048249","wikidata":"https://www.wikidata.org/wiki/Q37437","display_name":"Syntax","level":2,"score":0.5559999942779541},{"id":"https://openalex.org/C2780012671","wikidata":"https://www.wikidata.org/wiki/Q1343870","display_name":"Corollary","level":2,"score":0.42989999055862427},{"id":"https://openalex.org/C145912823","wikidata":"https://www.wikidata.org/wiki/Q113558","display_name":"Dynamics (music)","level":2,"score":0.42829999327659607},{"id":"https://openalex.org/C98045186","wikidata":"https://www.wikidata.org/wiki/Q205663","display_name":"Process (computing)","level":2,"score":0.42419999837875366},{"id":"https://openalex.org/C51632099","wikidata":"https://www.wikidata.org/wiki/Q3985153","display_name":"Training set","level":2,"score":0.4016000032424927},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.3693999946117401},{"id":"https://openalex.org/C153180895","wikidata":"https://www.wikidata.org/wiki/Q7148389","display_name":"Pattern recognition (psychology)","level":2,"score":0.36719998717308044},{"id":"https://openalex.org/C77405623","wikidata":"https://www.wikidata.org/wiki/Q598451","display_name":"System dynamics","level":2,"score":0.35760000348091125},{"id":"https://openalex.org/C80444323","wikidata":"https://www.wikidata.org/wiki/Q2878974","display_name":"Theoretical computer science","level":1,"score":0.35370001196861267},{"id":"https://openalex.org/C83665646","wikidata":"https://www.wikidata.org/wiki/Q42139305","display_name":"Feature vector","level":2,"score":0.35260000824928284},{"id":"https://openalex.org/C48209547","wikidata":"https://www.wikidata.org/wiki/Q1331104","display_name":"Controllability","level":2,"score":0.3472000062465668},{"id":"https://openalex.org/C11413529","wikidata":"https://www.wikidata.org/wiki/Q8366","display_name":"Algorithm","level":1,"score":0.33390000462532043},{"id":"https://openalex.org/C59404180","wikidata":"https://www.wikidata.org/wiki/Q17013334","display_name":"Feature learning","level":2,"score":0.32839998602867126},{"id":"https://openalex.org/C195324797","wikidata":"https://www.wikidata.org/wiki/Q33742","display_name":"Natural language","level":2,"score":0.311599999666214},{"id":"https://openalex.org/C2781067378","wikidata":"https://www.wikidata.org/wiki/Q17027399","display_name":"Interpretability","level":2,"score":0.30489999055862427},{"id":"https://openalex.org/C34736171","wikidata":"https://www.wikidata.org/wiki/Q918333","display_name":"Preprocessor","level":2,"score":0.304500013589859},{"id":"https://openalex.org/C204321447","wikidata":"https://www.wikidata.org/wiki/Q30642","display_name":"Natural language processing","level":1,"score":0.29649999737739563},{"id":"https://openalex.org/C52622490","wikidata":"https://www.wikidata.org/wiki/Q1026626","display_name":"Feature extraction","level":2,"score":0.28690001368522644},{"id":"https://openalex.org/C2777211547","wikidata":"https://www.wikidata.org/wiki/Q17141490","display_name":"Training (meteorology)","level":2,"score":0.25949999690055847}],"mesh":[],"locations_count":2,"locations":[{"id":"pmh:oai:arXiv.org:2502.20681","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2502.20681","pdf_url":"https://arxiv.org/pdf/2502.20681","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.2502.20681","doi":"https://doi.org/10.48550/arxiv.2502.20681","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2502.20681","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":null,"license_id":null,"version":null,"is_accepted":false,"is_published":null,"raw_source_name":null,"raw_type":"Preprint"}],"best_oa_location":{"id":"pmh:oai:arXiv.org:2502.20681","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2502.20681","pdf_url":"https://arxiv.org/pdf/2502.20681","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"},"sustainable_development_goals":[],"sustainable_development_goals_aurora":[{"id":"https://metadata.un.org/sdg/4","score":0.5922993421554565,"display_name":"Quality education"}],"study_designs":[],"awards":[{"id":"https://openalex.org/G8582114212","display_name":null,"funder_award_id":"2024YFE0203200","funder_id":"https://openalex.org/F4320335777","funder_display_name":"National Key Research and Development Program of China"}],"funders":[{"id":"https://openalex.org/F4320321001","display_name":"National Natural Science Foundation of China","ror":"https://ror.org/01h0zpd94"},{"id":"https://openalex.org/F4320322499","display_name":"Renmin University of China","ror":"https://ror.org/041pakw92"},{"id":"https://openalex.org/F4320335777","display_name":"National Key Research and Development Program of China","ror":null}],"has_content":{"grobid_xml":true,"pdf":true},"content_urls":{"pdf":"https://content.openalex.org/works/W6947817954.pdf","grobid_xml":"https://content.openalex.org/works/W6947817954.grobid-xml"},"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Transformers":[0],"may":[1],"exhibit":[2],"two-stage":[3,41,64,139,155],"training":[4,9,14,65],"dynamics":[5,66,77],"during":[6],"the":[7,17,20,63,74,80,132,161],"real-world":[8],"process.":[10],"For":[11],"instance,":[12],"when":[13],"GPT-2":[15],"on":[16],"Counterfact":[18],"dataset,":[19],"answers":[21],"progress":[22],"from":[23],"syntactically":[24,27],"incorrect":[25],"to":[26,29,48,160],"correct":[28],"semantically":[30],"correct.":[31],"However,":[32],"existing":[33],"theoretical":[34,146],"analyses":[35],"hardly":[36],"account":[37],"for":[38],"this":[39,57,130,145],"feature-level":[40,138],"phenomenon,":[42],"which":[43],"could":[44],"be":[45],"conceptually":[46],"attributed":[47],"disentangled":[49,82],"two-type":[50,83],"features":[51],"like":[52],"syntax":[53,116],"and":[54,100,117,119,123],"semantics.":[55],"In":[56],"paper,":[58],"we":[59,72],"theoretically":[60],"demonstrate":[61],"how":[62],"potentially":[67],"occur":[68],"in":[69,89,110,142],"transformers.":[70],"Specifically,":[71],"analyze":[73],"feature":[75,84,106],"learning":[76],"induced":[78],"by":[79],"aforementioned":[81],"structure,":[85],"grounding":[86],"our":[87,127],"analysis":[88],"a":[90,137,154],"simplified":[91],"yet":[92],"illustrative":[93],"setting":[94],"that":[95,152],"comprises":[96],"normalized":[97],"ReLU":[98],"self-attention":[99],"structured":[101],"data.":[102],"Such":[103],"disentanglement":[104],"of":[105,164],"structure":[107],"is":[108,131,157],"general":[109],"practice,":[111],"e.g.,":[112],"natural":[113],"languages":[114],"contain":[115,121],"semantics,":[118],"proteins":[120],"primary":[122],"secondary":[124],"structures.":[125],"To":[126],"best":[128],"knowledge,":[129],"first":[133],"rigorous":[134],"result":[135],"regarding":[136],"optimization":[140],"process":[141,156],"transformers":[143],"within":[144],"framework.":[147],"A":[148],"corollary":[149],"further":[150],"indicates":[151],"such":[153],"closely":[158],"related":[159],"spectral":[162],"properties":[163],"attention":[165],"weights.":[166]},"counts_by_year":[],"updated_date":"2026-10-15T10:00:46.086208","created_date":"2025-10-10T00:00:00"}