{"id":"https://openalex.org/W4416635575","doi":"https://doi.org/10.52202/085713-5320","title":"Superposition Yields Robust Neural Scaling","display_name":"Superposition Yields Robust Neural Scaling","publication_year":2025,"publication_date":"2025-01-01","ids":{"openalex":"https://openalex.org/W4416635575","doi":"https://doi.org/10.52202/085713-5320"},"language":"en","primary_location":{"id":"doi:10.52202/085713-5320","doi":"https://doi.org/10.52202/085713-5320","is_oa":false,"landing_page_url":"https://doi.org/10.52202/085713-5320","pdf_url":null,"source":{"id":"https://openalex.org/S4363606243","display_name":"neural information processing systems","issn_l":"1049-5258","issn":["1049-5258"],"is_oa":false,"is_in_doaj":false,"is_core":true,"listed_in":["cwts-core","norway-2"],"host_organization":null,"host_organization_name":null,"host_organization_lineage":[],"host_organization_lineage_names":[],"type":"conference"},"license":null,"license_id":null,"version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Advances in Neural Information Processing Systems 38","raw_type":"proceedings-article"},"volume":null,"type":"conference-paper","indexed_in":["arxiv","crossref","datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://arxiv.org/pdf/2505.10465","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":null,"display_name":"Yizhou Liu","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I63966007","display_name":"Massachusetts Institute of Technology","ror":"https://ror.org/042nb2s44","country_code":"US","type":"education","lineage":["https://openalex.org/I63966007"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Yizhou Liu","raw_affiliation_strings":["Massachusetts Institute of Technology"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Massachusetts Institute of Technology","institution_ids":["https://openalex.org/I63966007"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5101924541","display_name":"Ziming Liu","orcid":"https://orcid.org/0000-0003-3581-7494","observed_orcids":["https://orcid.org/0000-0003-3581-7494","https://orcid.org/0000-0002-7090-8853"]},"institutions":[{"id":"https://openalex.org/I1326498283","display_name":"Institut national de recherche en sciences et technologies du num\u00e9rique","ror":"https://ror.org/02kvxyf05","country_code":"FR","type":"government","lineage":["https://openalex.org/I1326498283"]}],"countries":["FR"],"is_corresponding":false,"raw_author_name":"Ziming Liu","raw_affiliation_strings":["INRIA"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"INRIA","institution_ids":["https://openalex.org/I1326498283"]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5003202779","display_name":"Jeff Gore","orcid":"https://orcid.org/0000-0003-4583-8555","observed_orcids":["https://orcid.org/0000-0003-4583-8555"]},"institutions":[{"id":"https://openalex.org/I63966007","display_name":"Massachusetts Institute of Technology","ror":"https://ror.org/042nb2s44","country_code":"US","type":"education","lineage":["https://openalex.org/I63966007"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Jeff Gore","raw_affiliation_strings":["Massachusetts Institute of Technology"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Massachusetts Institute of Technology","institution_ids":["https://openalex.org/I63966007"]}]}],"institutions":[],"countries_distinct_count":2,"institutions_distinct_count":2,"corresponding_author_ids":[],"corresponding_institution_ids":[],"apc_list":null,"apc_paid":null,"fwci":0.9483,"has_fulltext":true,"cited_by_count":1,"citation_normalized_percentile":{"value":0.78193544,"is_in_top_1_percent":false,"is_in_top_10_percent":false},"cited_by_percentile_year":{"min":93,"max":96},"biblio":{"volume":null,"issue":null,"first_page":"176757","last_page":"176793"},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T10028","display_name":"Topic Modeling","score":0.9082000255584717,"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":0.9082000255584717,"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/T10181","display_name":"Natural Language Processing Techniques","score":0.01679999940097332,"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":0.01679999940097332,"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/neural-scaling-laws","display_name":"neural scaling laws","score":0.9330000281333923},{"id":"https://openalex.org/keywords/large-language-models","display_name":"large language models","score":0.9549999833106995},{"id":"https://openalex.org/keywords/loss-scaling","display_name":"loss scaling","score":0.7319999933242798},{"id":"https://openalex.org/keywords/model-size","display_name":"model size","score":0.6320000290870667},{"id":"https://openalex.org/keywords/model-dimension","display_name":"model dimension","score":0.6890000104904175},{"id":"https://openalex.org/keywords/weight-decay","display_name":"weight decay","score":0.8330000042915344},{"id":"https://openalex.org/keywords/power-law-distribution","display_name":"power-law distribution","score":0.6380000114440918},{"id":"https://openalex.org/keywords/representational-geometry","display_name":"representational geometry","score":0.3919999897480011}],"concepts":[{"id":"https://openalex.org/C27753989","wikidata":"https://www.wikidata.org/wiki/Q284885","display_name":"Superposition principle","level":2,"score":0.7865999937057495},{"id":"https://openalex.org/C99844830","wikidata":"https://www.wikidata.org/wiki/Q102441924","display_name":"Scaling","level":2,"score":0.7718999981880188},{"id":"https://openalex.org/C2776359362","wikidata":"https://www.wikidata.org/wiki/Q2145286","display_name":"Representation (politics)","level":3,"score":0.6603999733924866},{"id":"https://openalex.org/C50644808","wikidata":"https://www.wikidata.org/wiki/Q192776","display_name":"Artificial neural network","level":2,"score":0.49570000171661377},{"id":"https://openalex.org/C33676613","wikidata":"https://www.wikidata.org/wiki/Q13415176","display_name":"Dimension (graph theory)","level":2,"score":0.492900013923645},{"id":"https://openalex.org/C2776401178","wikidata":"https://www.wikidata.org/wiki/Q12050496","display_name":"Feature (linguistics)","level":2,"score":0.47609999775886536},{"id":"https://openalex.org/C121864883","wikidata":"https://www.wikidata.org/wiki/Q677916","display_name":"Statistical physics","level":1,"score":0.47189998626708984},{"id":"https://openalex.org/C87040749","wikidata":"https://www.wikidata.org/wiki/Q428971","display_name":"Power law","level":2,"score":0.4512999951839447},{"id":"https://openalex.org/C33923547","wikidata":"https://www.wikidata.org/wiki/Q395","display_name":"Mathematics","level":0,"score":0.38589999079704285},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.3564000129699707},{"id":"https://openalex.org/C163258240","wikidata":"https://www.wikidata.org/wiki/Q25342","display_name":"Power (physics)","level":2,"score":0.35359999537467957},{"id":"https://openalex.org/C61797465","wikidata":"https://www.wikidata.org/wiki/Q1188986","display_name":"Term (time)","level":2,"score":0.3497999906539917},{"id":"https://openalex.org/C2778755073","wikidata":"https://www.wikidata.org/wiki/Q10858537","display_name":"Scale (ratio)","level":2,"score":0.3190999925136566},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.2976999878883362},{"id":"https://openalex.org/C180205008","wikidata":"https://www.wikidata.org/wiki/Q159190","display_name":"Amplitude","level":2,"score":0.29580000042915344},{"id":"https://openalex.org/C11413529","wikidata":"https://www.wikidata.org/wiki/Q8366","display_name":"Algorithm","level":1,"score":0.29280000925064087},{"id":"https://openalex.org/C121332964","wikidata":"https://www.wikidata.org/wiki/Q413","display_name":"Physics","level":0,"score":0.2685999870300293},{"id":"https://openalex.org/C2777212361","wikidata":"https://www.wikidata.org/wiki/Q5127848","display_name":"Class (philosophy)","level":2,"score":0.25200000405311584}],"mesh":[],"locations_count":3,"locations":[{"id":"doi:10.52202/085713-5320","doi":"https://doi.org/10.52202/085713-5320","is_oa":false,"landing_page_url":"https://doi.org/10.52202/085713-5320","pdf_url":null,"source":{"id":"https://openalex.org/S4363606243","display_name":"neural information processing systems","issn_l":"1049-5258","issn":["1049-5258"],"is_oa":false,"is_in_doaj":false,"is_core":true,"listed_in":["cwts-core","norway-2"],"host_organization":null,"host_organization_name":null,"host_organization_lineage":[],"host_organization_lineage_names":[],"type":"conference"},"license":null,"license_id":null,"version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Advances in Neural Information Processing Systems 38","raw_type":"proceedings-article"},{"id":"pmh:oai:arXiv.org:2505.10465","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2505.10465","pdf_url":"https://arxiv.org/pdf/2505.10465","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.2505.10465","doi":"https://doi.org/10.48550/arxiv.2505.10465","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2505.10465","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:2505.10465","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2505.10465","pdf_url":"https://arxiv.org/pdf/2505.10465","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":[{"id":"https://openalex.org/F4320306151","display_name":"Alfred P. Sloan Foundation","ror":"https://ror.org/052csg198"}],"has_content":{"grobid_xml":true,"pdf":true},"content_urls":{"pdf":"https://content.openalex.org/works/W4416635575.pdf","grobid_xml":"https://content.openalex.org/works/W4416635575.grobid-xml"},"referenced_works_count":0,"referenced_works":[],"related_works":["https://openalex.org/W7214979933","https://openalex.org/W4360892090","https://openalex.org/W7214737100","https://openalex.org/W7127887711","https://openalex.org/W7214537689","https://openalex.org/W7162298793","https://openalex.org/W4407759068","https://openalex.org/W7165905290","https://openalex.org/W4415084622","https://openalex.org/W4307929880"],"abstract_inverted_index":{"The":[0],"success":[1],"of":[2,20,76,124,174],"today's":[3],"large":[4],"language":[5],"models":[6,14],"(LLMs)":[7],"depends":[8],"on":[9,64],"the":[10,18,74,93,112,141,155],"observation":[11],"that":[12,25,39,43,136,154],"larger":[13],"perform":[15],"better.":[16],"However,":[17],"origin":[19],"this":[21,163],"neural":[22,61,175,184],"scaling":[23,148,157,176,185],"law,":[24],"loss":[26,58,84,94,113,147],"decreases":[27],"as":[28,170],"a":[29,54,96,121,171],"power":[30,97],"law":[31,98],"with":[32,86,117,150,162],"model":[33,87,118,151],"size,":[34],"remains":[35],"unclear.":[36],"We":[37,134],"propose":[38],"representation":[40,132,168],"superposition,":[41,77,111],"meaning":[42],"LLMs":[44,138],"represent":[45],"more":[46],"features":[47],"than":[48],"they":[49,192],"have":[50,146],"dimensions,":[51],"can":[52,187],"be":[53,188],"key":[55],"contributor":[56],"to":[57,72,80,128],"and":[59,145,153,190],"cause":[60],"scaling.":[62],"Based":[63],"Anthropic's":[65],"toy":[66],"model,":[67],"we":[68],"use":[69],"weight":[70],"decay":[71],"control":[73],"degree":[75],"allowing":[78],"us":[79],"systematically":[81],"study":[82],"how":[83],"scales":[85,115],"size.":[88],"When":[89],"superposition":[90,143,169],"is":[91],"weak,":[92],"follows":[95],"only":[99],"if":[100],"data":[101],"feature":[102],"frequencies":[103],"are":[104,159],"power-law":[105],"distributed.":[106],"In":[107],"contrast,":[108],"under":[109],"strong":[110,142],"generically":[114],"inversely":[116,149],"dimension":[119],"across":[120],"broad":[122],"class":[123],"frequency":[125],"distributions,":[126],"due":[127],"geometric":[129],"overlaps":[130],"between":[131],"vectors.":[133],"confirmed":[135],"open-sourced":[137],"operate":[139],"in":[140],"regime":[144],"dimension,":[152],"Chinchilla":[156],"laws":[158,186],"also":[160],"consistent":[161],"behavior.":[164],"Our":[165],"results":[166],"identify":[167],"central":[172],"driver":[173],"laws,":[177],"providing":[178],"insights":[179],"into":[180],"questions":[181],"like":[182],"when":[183,191],"improved":[189],"will":[193],"break":[194],"down.":[195]},"counts_by_year":[{"year":2026,"cited_by_count":1}],"updated_date":"2026-10-16T09:59:22.235173","created_date":"2025-10-10T00:00:00"}