{"id":"https://openalex.org/W7140124962","doi":"https://doi.org/10.48550/arxiv.2603.19348","title":"Anatomical Heterogeneity in Transformer Language Models","display_name":"Anatomical Heterogeneity in Transformer Language Models","publication_year":2026,"publication_date":"2026-03-19","ids":{"openalex":"https://openalex.org/W7140124962","doi":"https://doi.org/10.48550/arxiv.2603.19348"},"language":"en","primary_location":{"id":"pmh:oai:arXiv.org:2603.19348","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2603.19348","pdf_url":"https://arxiv.org/pdf/2603.19348","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","is_traditional":true,"indexed_in":["arxiv","datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://arxiv.org/pdf/2603.19348","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5130405324","display_name":"Tomasz Wietrzykowski","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":true,"raw_author_name":"Wietrzykowski, Tomasz","raw_affiliation_strings":["Independent Researcher , Wroclaw , Poland"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Independent Researcher , Wroclaw , Poland","institution_ids":[]}]}],"institutions":[],"countries_distinct_count":0,"institutions_distinct_count":0,"corresponding_author_ids":["https://openalex.org/A5130405324"],"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":false,"primary_topic":{"id":"https://openalex.org/T10028","display_name":"Topic Modeling","score":0.8432999849319458,"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":0.8432999849319458,"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/T10036","display_name":"Advanced Neural Network Applications","score":0.052400000393390656,"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/T10181","display_name":"Natural Language Processing Techniques","score":0.051100000739097595,"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-based-language-models","display_name":"transformer language models","score":0.8939999938011169},{"id":"https://openalex.org/keywords/layer-importance","display_name":"layer importance","score":0.6710000038146973},{"id":"https://openalex.org/keywords/catastrophic-forgetting","display_name":"catastrophic forgetting","score":0.5210000276565552},{"id":"https://openalex.org/keywords/training-budget-allocation","display_name":"training budget allocation","score":0.5979999899864197},{"id":"https://openalex.org/keywords/causal-language-models","display_name":"causal language models","score":0.6859999895095825}],"concepts":[{"id":"https://openalex.org/C197640229","wikidata":"https://www.wikidata.org/wiki/Q2534066","display_name":"Predictability","level":2,"score":0.7287999987602234},{"id":"https://openalex.org/C66322947","wikidata":"https://www.wikidata.org/wiki/Q11658","display_name":"Transformer","level":3,"score":0.6244999766349792},{"id":"https://openalex.org/C137293760","wikidata":"https://www.wikidata.org/wiki/Q3621696","display_name":"Language model","level":2,"score":0.6150000095367432},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.5351999998092651},{"id":"https://openalex.org/C99844830","wikidata":"https://www.wikidata.org/wiki/Q102441924","display_name":"Scaling","level":2,"score":0.5257999897003174},{"id":"https://openalex.org/C158622935","wikidata":"https://www.wikidata.org/wiki/Q660848","display_name":"Nonlinear system","level":2,"score":0.5246000289916992},{"id":"https://openalex.org/C11413529","wikidata":"https://www.wikidata.org/wiki/Q8366","display_name":"Algorithm","level":1,"score":0.3668000102043152},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.35280001163482666},{"id":"https://openalex.org/C33923547","wikidata":"https://www.wikidata.org/wiki/Q395","display_name":"Mathematics","level":0,"score":0.33469998836517334},{"id":"https://openalex.org/C2779227376","wikidata":"https://www.wikidata.org/wiki/Q6505497","display_name":"Layer (electronics)","level":2,"score":0.3215000033378601},{"id":"https://openalex.org/C2984755018","wikidata":"https://www.wikidata.org/wiki/Q17118374","display_name":"Nonlinear model","level":3,"score":0.2867000102996826},{"id":"https://openalex.org/C47446073","wikidata":"https://www.wikidata.org/wiki/Q5165890","display_name":"Control theory (sociology)","level":3,"score":0.26100000739097595},{"id":"https://openalex.org/C51632099","wikidata":"https://www.wikidata.org/wiki/Q3985153","display_name":"Training set","level":2,"score":0.26100000739097595}],"mesh":[],"locations_count":2,"locations":[{"id":"pmh:oai:arXiv.org:2603.19348","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2603.19348","pdf_url":"https://arxiv.org/pdf/2603.19348","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.2603.19348","doi":"https://doi.org/10.48550/arxiv.2603.19348","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2603.19348","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:2603.19348","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2603.19348","pdf_url":"https://arxiv.org/pdf/2603.19348","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":[],"funders":[{"id":"https://openalex.org/F4320309480","display_name":"Nvidia","ror":"https://ror.org/03jdj4y14"}],"has_content":{"grobid_xml":true,"pdf":true},"content_urls":{"pdf":"https://content.openalex.org/works/W7140124962.pdf","grobid_xml":"https://content.openalex.org/works/W7140124962.grobid-xml"},"referenced_works_count":0,"referenced_works":[],"related_works":["https://openalex.org/W7223439878","https://openalex.org/W4409362497","https://openalex.org/W7163738417","https://openalex.org/W3023662336","https://openalex.org/W4389524424","https://openalex.org/W7140195227","https://openalex.org/W7167476553","https://openalex.org/W7165659024","https://openalex.org/W7140193867","https://openalex.org/W7207058747"],"abstract_inverted_index":{"Current":[0],"transformer":[1],"language":[2,30],"models":[3],"are":[4],"trained":[5],"with":[6,64,113],"uniform":[7,158],"computational":[8],"budgets":[9],"across":[10],"all":[11],"layers,":[12],"implicitly":[13],"assuming":[14],"layer":[15,114,142],"homogeneity.":[16],"We":[17,49],"challenge":[18],"this":[19],"assumption":[20],"through":[21],"empirical":[22],"analysis":[23],"of":[24],"SmolLM2-135M,":[25],"a":[26,65,88,92],"30-layer,":[27],"135M-parameter":[28],"causal":[29],"model,":[31],"using":[32],"five":[33,131],"diagnostic":[34],"metrics:":[35],"weight":[36,43,122],"predictability":[37],"(R2),":[38],"ablation":[39],"degradation,":[40],"recovery":[41],"speed,":[42],"manipulation":[44,133],"robustness,":[45],"and":[46],"structural":[47],"analysis.":[48],"find":[50],"profound":[51],"anatomical":[52],"heterogeneity:":[53],"(1)":[54],"Layer":[55,85],"weights":[56,75],"follow":[57],"strong":[58],"mathematical":[59],"regularity":[60],"(R2":[61],"=":[62,125],"0.91)":[63],"universal":[66],"oscillatory":[67],"delta":[68],"pattern":[69],"(correlation":[70],"~=":[71],"-0.50),":[72],"yet":[73],"predicted":[74],"cause":[76],"catastrophic":[77],"failure":[78],"due":[79],"to":[80,97,101],"nonlinear":[81],"error":[82],"accumulation.":[83],"(2)":[84],"importance":[86],"spans":[87],"10^7":[89],"range,":[90],"from":[91],"critical":[93],"core":[94],"(L8-11,":[95],"up":[96],"+63,419%":[98],"PPL":[99],"degradation)":[100],"anti-layers":[102],"(L14,":[103],"L17)":[104],"whose":[105],"removal":[106],"improves":[107],"performance.":[108],"(3)":[109],"Recovery":[110],"speed":[111],"correlates":[112],"importance,":[115,143],"indicating":[116],"differential":[117],"training":[118,159],"requirements.":[119],"(4)":[120],"Only":[121],"scaling":[123],"(alpha":[124],"0.9)":[126],"preserves":[127],"model":[128],"quality":[129],"among":[130],"tested":[132],"strategies.":[134],"(5)":[135],"Growth":[136],"Transformer":[137],"Training,":[138],"allocating":[139],"budget":[140],"by":[141],"achieves":[144],"~54%":[145],"cost":[146],"reduction.":[147],"A":[148],"proof-of-concept":[149],"experiment":[150],"confirms":[151],"this:":[152],"4.7x":[153],"lower":[154],"validation":[155],"loss":[156],"than":[157],"at":[160],"identical":[161],"parameter":[162],"count,":[163],"while":[164],"being":[165],"13%":[166],"faster.":[167]},"counts_by_year":[],"updated_date":"2026-10-28T10:33:24.094632","created_date":"2026-03-24T00:00:00"}