{"id":"https://openalex.org/W4417302244","doi":"https://doi.org/10.48550/arxiv.2505.12822","title":"Emergent Specialization: Rare Token Neurons in Language Models","display_name":"Emergent Specialization: Rare Token Neurons in Language Models","publication_year":2025,"publication_date":"2025-05-19","ids":{"openalex":"https://openalex.org/W4417302244","doi":"https://doi.org/10.48550/arxiv.2505.12822"},"language":"en","primary_location":{"id":"pmh:oai:arXiv.org:2505.12822","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2505.12822","pdf_url":"https://arxiv.org/pdf/2505.12822","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","indexed_in":["arxiv","datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://arxiv.org/pdf/2505.12822","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":null,"display_name":"Liu, Jing","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I1294671590","display_name":"Centre National de la Recherche Scientifique","ror":"https://ror.org/02feahw73","country_code":"FR","type":"government","lineage":["https://openalex.org/I1294671590"]},{"id":"https://openalex.org/I2746051580","display_name":"Universit\u00e9 Paris Sciences et Lettres","ror":"https://ror.org/013cjyk83","country_code":"FR","type":"education","lineage":["https://openalex.org/I2746051580"]},{"id":"https://openalex.org/I29607241","display_name":"\u00c9cole Normale Sup\u00e9rieure - PSL","ror":"https://ror.org/05a0dhs15","country_code":"FR","type":"other","lineage":["https://openalex.org/I2746051580","https://openalex.org/I29607241"]},{"id":"https://openalex.org/I90669466","display_name":"\u00c9cole des hautes \u00e9tudes en sciences sociales","ror":"https://ror.org/02d9dg697","country_code":"FR","type":"facility","lineage":["https://openalex.org/I90669466"]}],"countries":["FR"],"is_corresponding":false,"raw_author_name":"Liu, Jing","raw_affiliation_strings":["ENS , Universit\u00e9 PSL , EHESS , CNRS Paris , France"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"ENS , Universit\u00e9 PSL , EHESS , CNRS Paris , France","institution_ids":["https://openalex.org/I1294671590","https://openalex.org/I2746051580","https://openalex.org/I29607241","https://openalex.org/I90669466"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5036278941","display_name":"H. Wang","orcid":"https://orcid.org/0009-0003-8329-0145","observed_orcids":["https://orcid.org/0009-0003-8329-0145"]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Wang, Haozheng","raw_affiliation_strings":["DT Master Carbon","Independant Researcher Paris , France"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"DT Master Carbon","institution_ids":[]},{"raw_affiliation_string":"Independant Researcher Paris , France","institution_ids":[]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5101753750","display_name":"Yueheng Li","orcid":"https://orcid.org/0000-0003-1327-3334","observed_orcids":["https://orcid.org/0000-0003-1327-3334"]},"institutions":[{"id":"https://openalex.org/I1294671590","display_name":"Centre National de la Recherche Scientifique","ror":"https://ror.org/02feahw73","country_code":"FR","type":"government","lineage":["https://openalex.org/I1294671590"]},{"id":"https://openalex.org/I204730241","display_name":"Universit\u00e9 Paris Cit\u00e9","ror":"https://ror.org/05f82e368","country_code":"FR","type":"education","lineage":["https://openalex.org/I204730241"]},{"id":"https://openalex.org/I29607241","display_name":"\u00c9cole Normale Sup\u00e9rieure - PSL","ror":"https://ror.org/05a0dhs15","country_code":"FR","type":"other","lineage":["https://openalex.org/I2746051580","https://openalex.org/I29607241"]},{"id":"https://openalex.org/I39804081","display_name":"Sorbonne Universit\u00e9","ror":"https://ror.org/02en5vm52","country_code":"FR","type":"education","lineage":["https://openalex.org/I39804081"]},{"id":"https://openalex.org/I4210091437","display_name":"Sorbonne Paris Cit\u00e9","ror":"https://ror.org/001z21q04","country_code":"FR","type":"other","lineage":["https://openalex.org/I204730241","https://openalex.org/I4210091437"]},{"id":"https://openalex.org/I4210103330","display_name":"\u00c9cole Normale Sup\u00e9rieure","ror":"https://ror.org/01dp7jr64","country_code":"BI","type":"education","lineage":["https://openalex.org/I4210103330"]},{"id":"https://openalex.org/I4210129651","display_name":"Laboratoire de Physique de l'Ecole Normale Sup\u00e9rieure","ror":"https://ror.org/03a26mh11","country_code":"FR","type":"facility","lineage":["https://openalex.org/I1294671590","https://openalex.org/I204730241","https://openalex.org/I2746051580","https://openalex.org/I2800004676","https://openalex.org/I29607241","https://openalex.org/I39804081","https://openalex.org/I4210129651"]}],"countries":["BI","FR"],"is_corresponding":false,"raw_author_name":"Li, Yueheng","raw_affiliation_strings":["Laboratoire de Physique de l'\u00c9cole Normale Sup\u00e9rieure , ENS , Universit\u00e9 PSL , CNRS , Sorbonne Universit\u00e9 , Universit\u00e9 Paris Cit\u00e9 Paris , France","Sorbonne Universit\u00e9 , \u00c9cole Normale Sup\u00e9rieure , CNRS , Laboratoire de Physique (LPENS)"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Laboratoire de Physique de l'\u00c9cole Normale Sup\u00e9rieure , ENS , Universit\u00e9 PSL , CNRS , Sorbonne Universit\u00e9 , Universit\u00e9 Paris Cit\u00e9 Paris , France","institution_ids":["https://openalex.org/I1294671590","https://openalex.org/I204730241","https://openalex.org/I29607241","https://openalex.org/I39804081","https://openalex.org/I4210091437","https://openalex.org/I4210103330"]},{"raw_affiliation_string":"Sorbonne Universit\u00e9 , \u00c9cole Normale Sup\u00e9rieure , CNRS , Laboratoire de Physique (LPENS)","institution_ids":["https://openalex.org/I1294671590","https://openalex.org/I39804081","https://openalex.org/I4210129651"]}]}],"institutions":[],"countries_distinct_count":2,"institutions_distinct_count":9,"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":false,"primary_topic":{"id":"https://openalex.org/T10028","display_name":"Topic Modeling","score":0.8647000193595886,"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":0.8647000193595886,"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":0.034699998795986176,"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":0.029100000858306885,"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/large-language-models","display_name":"large language models","score":0.871999979019165},{"id":"https://openalex.org/keywords/co-activation","display_name":"CO activation","score":0.6129999756813049},{"id":"https://openalex.org/keywords/statistical-mechanics-of-neural-networks","display_name":"statistical mechanics of neural networks","score":0.8190000057220459},{"id":"https://openalex.org/keywords/functional-specialization","display_name":"functional specialization","score":0.38600000739097595}],"concepts":[{"id":"https://openalex.org/C2780186347","wikidata":"https://www.wikidata.org/wiki/Q11414","display_name":"Subnetwork","level":2,"score":0.8043000102043152},{"id":"https://openalex.org/C48145219","wikidata":"https://www.wikidata.org/wiki/Q1335365","display_name":"Security token","level":2,"score":0.7820000052452087},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.6309999823570251},{"id":"https://openalex.org/C137293760","wikidata":"https://www.wikidata.org/wiki/Q3621696","display_name":"Language model","level":2,"score":0.5134000182151794},{"id":"https://openalex.org/C66882249","wikidata":"https://www.wikidata.org/wiki/Q169336","display_name":"Homogeneous","level":2,"score":0.4691999852657318},{"id":"https://openalex.org/C89611455","wikidata":"https://www.wikidata.org/wiki/Q6804646","display_name":"Mechanism (biology)","level":2,"score":0.45350000262260437},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.3984000086784363},{"id":"https://openalex.org/C48103436","wikidata":"https://www.wikidata.org/wiki/Q599031","display_name":"State (computer science)","level":2,"score":0.3871000111103058},{"id":"https://openalex.org/C169760540","wikidata":"https://www.wikidata.org/wiki/Q207011","display_name":"Neuroscience","level":1,"score":0.3675000071525574},{"id":"https://openalex.org/C2777317252","wikidata":"https://www.wikidata.org/wiki/Q18393516","display_name":"Rare events","level":2,"score":0.3659000098705292},{"id":"https://openalex.org/C12426560","wikidata":"https://www.wikidata.org/wiki/Q189569","display_name":"Basis (linear algebra)","level":2,"score":0.36559998989105225},{"id":"https://openalex.org/C204321447","wikidata":"https://www.wikidata.org/wiki/Q30642","display_name":"Natural language processing","level":1,"score":0.32199999690055847},{"id":"https://openalex.org/C15744967","wikidata":"https://www.wikidata.org/wiki/Q9418","display_name":"Psychology","level":0,"score":0.30889999866485596},{"id":"https://openalex.org/C114289077","wikidata":"https://www.wikidata.org/wiki/Q3284399","display_name":"Statistical model","level":2,"score":0.30630001425743103},{"id":"https://openalex.org/C188147891","wikidata":"https://www.wikidata.org/wiki/Q147638","display_name":"Cognitive science","level":1,"score":0.29649999737739563},{"id":"https://openalex.org/C46312422","wikidata":"https://www.wikidata.org/wiki/Q11024","display_name":"Communication","level":1,"score":0.29339998960494995},{"id":"https://openalex.org/C179603123","wikidata":"https://www.wikidata.org/wiki/Q1941921","display_name":"Modeling language","level":3,"score":0.275299996137619},{"id":"https://openalex.org/C125411270","wikidata":"https://www.wikidata.org/wiki/Q18653","display_name":"Encoding (memory)","level":2,"score":0.2732999920845032},{"id":"https://openalex.org/C66024118","wikidata":"https://www.wikidata.org/wiki/Q1122506","display_name":"Computational model","level":2,"score":0.272599995136261},{"id":"https://openalex.org/C2993724205","wikidata":"https://www.wikidata.org/wiki/Q315","display_name":"Human language","level":2,"score":0.27149999141693115},{"id":"https://openalex.org/C98045186","wikidata":"https://www.wikidata.org/wiki/Q205663","display_name":"Process (computing)","level":2,"score":0.2696000039577484},{"id":"https://openalex.org/C177264268","wikidata":"https://www.wikidata.org/wiki/Q1514741","display_name":"Set (abstract data type)","level":2,"score":0.26420000195503235},{"id":"https://openalex.org/C26760741","wikidata":"https://www.wikidata.org/wiki/Q160402","display_name":"Perception","level":2,"score":0.2632000148296356},{"id":"https://openalex.org/C94922259","wikidata":"https://www.wikidata.org/wiki/Q33215","display_name":"Constructed language","level":2,"score":0.2556000053882599},{"id":"https://openalex.org/C195324797","wikidata":"https://www.wikidata.org/wiki/Q33742","display_name":"Natural language","level":2,"score":0.2500999867916107}],"mesh":[],"locations_count":2,"locations":[{"id":"pmh:oai:arXiv.org:2505.12822","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2505.12822","pdf_url":"https://arxiv.org/pdf/2505.12822","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.2505.12822","doi":"https://doi.org/10.48550/arxiv.2505.12822","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2505.12822","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":null,"raw_source_name":null,"raw_type":"Preprint"}],"best_oa_location":{"id":"pmh:oai:arXiv.org:2505.12822","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2505.12822","pdf_url":"https://arxiv.org/pdf/2505.12822","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":[],"sustainable_development_goals_aurora":[],"study_designs":[],"awards":[],"funders":[],"has_content":{"pdf":true,"grobid_xml":true},"content_urls":{"pdf":"https://content.openalex.org/works/W4417302244.pdf","grobid_xml":"https://content.openalex.org/works/W4417302244.grobid-xml"},"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Large":[0],"language":[1,28],"models":[2],"struggle":[3],"with":[4,23,93,101],"representing":[5],"and":[6,39,46,57],"generating":[7],"rare":[8,32,36,80],"tokens":[9],"despite":[10],"their":[11,44],"importance":[12],"in":[13],"specialized":[14],"domains.":[15],"In":[16,76],"this":[17],"study,":[18],"we":[19],"identify":[20],"neuron":[21],"structures":[22],"exceptionally":[24],"strong":[25],"influence":[26],"on":[27],"model's":[29],"prediction":[30],"of":[31,104],"tokens,":[33],"termed":[34],"as":[35],"token":[37,81],"neurons,":[38],"investigate":[40],"the":[41,77,102],"mechanism":[42],"for":[43,113],"emergence":[45],"behavior.":[47],"These":[48],"neurons":[49,82],"exhibit":[50],"a":[51,67,72,84,109],"characteristic":[52],"three-phase":[53],"organization":[54],"(plateau,":[55],"power-law,":[56],"rapid":[58],"decay)":[59],"that":[60,87],"emerges":[61],"dynamically":[62],"during":[63],"training,":[64],"evolving":[65],"from":[66],"homogeneous":[68],"initial":[69],"state":[70],"to":[71],"functionally":[73],"differentiated":[74],"architecture.":[75],"activation":[78],"space,":[79],"form":[83],"coordinated":[85],"subnetwork":[86],"selectively":[88],"co-activates":[89],"while":[90],"avoiding":[91],"co-activation":[92],"other":[94],"neurons.":[95],"This":[96],"functional":[97],"specialization":[98],"potentially":[99],"correlates":[100],"development":[103],"heavy-tailed":[105],"weight":[106],"distributions,":[107],"suggesting":[108],"statistical":[110],"mechanical":[111],"basis":[112],"emergent":[114],"specialization.":[115]},"counts_by_year":[],"updated_date":"2026-10-22T18:32:59.640997","created_date":"2025-10-10T00:00:00"}