{"id":"https://openalex.org/W4416036519","doi":"https://doi.org/10.18653/v1/2025.emnlp-main.593","title":"Data Drives Unstable Hierarchical Generalization in LMs","display_name":"Data Drives Unstable Hierarchical Generalization in LMs","publication_year":2025,"publication_date":"2025-01-01","ids":{"openalex":"https://openalex.org/W4416036519","doi":"https://doi.org/10.18653/v1/2025.emnlp-main.593"},"language":null,"primary_location":{"id":"doi:10.18653/v1/2025.emnlp-main.593","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2025.emnlp-main.593","pdf_url":"https://aclanthology.org/2025.emnlp-main.593.pdf","source":null,"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing","raw_type":"proceedings-article"},"type":"article","indexed_in":["crossref"],"open_access":{"is_oa":true,"oa_status":"gold","oa_url":"https://aclanthology.org/2025.emnlp-main.593.pdf","any_repository_has_fulltext":null},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5103262975","display_name":"Tian Qin","orcid":"https://orcid.org/0000-0002-3466-6041"},"institutions":[{"id":"https://openalex.org/I2801851002","display_name":"Harvard University Press","ror":"https://ror.org/006v7bf86","country_code":"US","type":"other","lineage":["https://openalex.org/I136199984","https://openalex.org/I2801851002"]}],"countries":["US"],"is_corresponding":true,"raw_author_name":"Tian Qin","raw_affiliation_strings":["Harvard University Harvard University Harvard University & MSR"],"affiliations":[{"raw_affiliation_string":"Harvard University Harvard University Harvard University & MSR","institution_ids":["https://openalex.org/I2801851002"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5010251681","display_name":"Naomi Saphra","orcid":"https://orcid.org/0000-0001-8607-3515"},"institutions":[{"id":"https://openalex.org/I2801851002","display_name":"Harvard University Press","ror":"https://ror.org/006v7bf86","country_code":"US","type":"other","lineage":["https://openalex.org/I136199984","https://openalex.org/I2801851002"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Naomi Saphra","raw_affiliation_strings":["Harvard University Harvard University Harvard University & MSR"],"affiliations":[{"raw_affiliation_string":"Harvard University Harvard University Harvard University & MSR","institution_ids":["https://openalex.org/I2801851002"]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5110947683","display_name":"David Alvarez-Melis","orcid":null},"institutions":[{"id":"https://openalex.org/I2801851002","display_name":"Harvard University Press","ror":"https://ror.org/006v7bf86","country_code":"US","type":"other","lineage":["https://openalex.org/I136199984","https://openalex.org/I2801851002"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"David Alvarez-Melis","raw_affiliation_strings":["Harvard University Harvard University Harvard University & MSR"],"affiliations":[{"raw_affiliation_string":"Harvard University Harvard University Harvard University & MSR","institution_ids":["https://openalex.org/I2801851002"]}]}],"institutions":[],"countries_distinct_count":1,"institutions_distinct_count":3,"corresponding_author_ids":["https://openalex.org/A5103262975"],"corresponding_institution_ids":["https://openalex.org/I2801851002"],"apc_list":null,"apc_paid":null,"fwci":0.0,"has_fulltext":true,"cited_by_count":0,"citation_normalized_percentile":{"value":0.18397168,"is_in_top_1_percent":false,"is_in_top_10_percent":false},"cited_by_percentile_year":null,"biblio":{"volume":null,"issue":null,"first_page":"11733","last_page":"11751"},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T10181","display_name":"Natural Language Processing Techniques","score":0.0869000032544136,"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/T10181","display_name":"Natural Language Processing Techniques","score":0.0869000032544136,"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/T11902","display_name":"Intelligent Tutoring Systems and Adaptive Learning","score":0.08209999650716782,"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/T10636","display_name":"Innovative Teaching and Learning Methods","score":0.04010000079870224,"subfield":{"id":"https://openalex.org/subfields/3204","display_name":"Developmental and Educational Psychology"},"field":{"id":"https://openalex.org/fields/32","display_name":"Psychology"},"domain":{"id":"https://openalex.org/domains/2","display_name":"Social Sciences"}}],"keywords":[{"id":"https://openalex.org/keywords/generalization","display_name":"Generalization","score":0.5491999983787537},{"id":"https://openalex.org/keywords/stability","display_name":"Stability (learning theory)","score":0.40070000290870667},{"id":"https://openalex.org/keywords/set","display_name":"Set (abstract data type)","score":0.2987000048160553},{"id":"https://openalex.org/keywords/feature","display_name":"Feature (linguistics)","score":0.2687999904155731},{"id":"https://openalex.org/keywords/noise","display_name":"Noise (video)","score":0.24770000576972961}],"concepts":[{"id":"https://openalex.org/C177148314","wikidata":"https://www.wikidata.org/wiki/Q170084","display_name":"Generalization","level":2,"score":0.5491999983787537},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.5394999980926514},{"id":"https://openalex.org/C112972136","wikidata":"https://www.wikidata.org/wiki/Q7595718","display_name":"Stability (learning theory)","level":2,"score":0.40070000290870667},{"id":"https://openalex.org/C11413529","wikidata":"https://www.wikidata.org/wiki/Q8366","display_name":"Algorithm","level":1,"score":0.38580000400543213},{"id":"https://openalex.org/C33923547","wikidata":"https://www.wikidata.org/wiki/Q395","display_name":"Mathematics","level":0,"score":0.3763999938964844},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.36719998717308044},{"id":"https://openalex.org/C177264268","wikidata":"https://www.wikidata.org/wiki/Q1514741","display_name":"Set (abstract data type)","level":2,"score":0.2987000048160553},{"id":"https://openalex.org/C2776401178","wikidata":"https://www.wikidata.org/wiki/Q12050496","display_name":"Feature (linguistics)","level":2,"score":0.2687999904155731},{"id":"https://openalex.org/C99498987","wikidata":"https://www.wikidata.org/wiki/Q2210247","display_name":"Noise (video)","level":3,"score":0.24770000576972961},{"id":"https://openalex.org/C58489278","wikidata":"https://www.wikidata.org/wiki/Q1172284","display_name":"Data set","level":2,"score":0.23989999294281006}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.18653/v1/2025.emnlp-main.593","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2025.emnlp-main.593","pdf_url":"https://aclanthology.org/2025.emnlp-main.593.pdf","source":null,"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing","raw_type":"proceedings-article"}],"best_oa_location":{"id":"doi:10.18653/v1/2025.emnlp-main.593","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2025.emnlp-main.593","pdf_url":"https://aclanthology.org/2025.emnlp-main.593.pdf","source":null,"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing","raw_type":"proceedings-article"},"sustainable_development_goals":[],"awards":[],"funders":[],"has_content":{"pdf":true,"grobid_xml":true},"content_urls":{"pdf":"https://content.openalex.org/works/W4416036519.pdf","grobid_xml":"https://content.openalex.org/works/W4416036519.grobid-xml"},"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Early":[0],"in":[1,73,107,159],"training,":[2],"LMs":[3],"can":[4,167],"behave":[5],"like":[6],"n-gram":[7],"models,":[8],"but":[9],"eventually,":[10],"they":[11],"often":[12],"learn":[13],"treebased":[14],"syntactic":[15,57,127],"rules":[16,86],"and":[17,33,131,145,162],"generalize":[18,41],"hierarchically":[19,42],"out":[20],"of":[21,76,125,156],"distribution":[22],"(OOD).We":[23],"study":[24],"this":[25,59,111],"shift":[26],"using":[27],"controlled":[28],"grammar-learning":[29],"tasks:":[30],"question":[31],"formation":[32],"tense":[34],"inflection.We":[35],"find":[36,81],"a":[37,55,83],"model":[38,84],"learns":[39],"to":[40,87,169],"if":[43,50,93,100],"its":[44,94],"training":[45,95,109,157],"data":[46,62,69,96,114,122,158],"is":[47,97,139],"complex-in":[48],"particular,":[49,99],"it":[51,101],"includes":[52,102],"center-embedded":[53],"clauses,":[54],"special":[56],"structure.Under":[58],"definition,":[60,112],"complex":[61,68],"drives":[63],"hierarchical":[64,90],"rules,":[65],"while":[66],"less":[67,120],"encourages":[70],"shortcut":[71],"learning":[72,143],"the":[74,108,153],"form":[75,134],"n-gram-like":[77],"linear":[78],"rules.Furthermore,":[79],"we":[80],"that":[82],"uses":[85],"generalize,":[88],"whether":[89],"or":[91],"linear,":[92],"diverse-in":[98],"many":[103],"distinct":[104],"syntax":[105],"trees":[106],"set.Under":[110],"diverse":[113,121],"promotes":[115,123],"stable":[116],"rule":[117],"learning,":[118],"whereas":[119],"memorization":[124],"individual":[126],"sequences.Finally,":[128],"intermediate":[129,132],"diversity":[130],"complexity":[133],"an":[135],"unstable":[136,170],"regime,":[137],"which":[138],"characterized":[140],"by":[141],"oscillatory":[142],"dynamics":[144],"inconsistent":[146],"behaviors":[147],"across":[148],"random":[149],"seeds.These":[150],"results":[151],"highlight":[152],"central":[154],"role":[155],"shaping":[160],"generalization":[161],"explain":[163],"why":[164],"competing":[165],"strategies":[166],"lead":[168],"outcomes.":[171]},"counts_by_year":[],"updated_date":"2026-03-13T14:20:09.374765","created_date":"2025-11-08T00:00:00"}