{"id":"https://openalex.org/W4416036836","doi":"https://doi.org/10.18653/v1/2025.emnlp-main.251","title":"Through the Valley: Path to Effective Long CoT Training for Small Language Models","display_name":"Through the Valley: Path to Effective Long CoT Training for Small Language Models","publication_year":2025,"publication_date":"2025-01-01","ids":{"openalex":"https://openalex.org/W4416036836","doi":"https://doi.org/10.18653/v1/2025.emnlp-main.251"},"language":null,"primary_location":{"id":"doi:10.18653/v1/2025.emnlp-main.251","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2025.emnlp-main.251","pdf_url":"https://aclanthology.org/2025.emnlp-main.251.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.251.pdf","any_repository_has_fulltext":null},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5100684743","display_name":"Ruikun Luo","orcid":"https://orcid.org/0009-0008-6015-9992"},"institutions":[],"countries":[],"is_corresponding":true,"raw_author_name":"Renjie Luo","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5104174699","display_name":"Jiaxi Li","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Jiaxi Li","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5101570817","display_name":"Chen Huang","orcid":"https://orcid.org/0000-0003-1626-8549"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Chen Huang","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5113599845","display_name":"Wei Lu","orcid":"https://orcid.org/0000-0002-0929-7416"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Wei Lu","raw_affiliation_strings":[],"affiliations":[]}],"institutions":[],"countries_distinct_count":0,"institutions_distinct_count":4,"corresponding_author_ids":["https://openalex.org/A5100684743"],"corresponding_institution_ids":[],"apc_list":null,"apc_paid":null,"fwci":0.0,"has_fulltext":true,"cited_by_count":0,"citation_normalized_percentile":{"value":0.17416542,"is_in_top_1_percent":false,"is_in_top_10_percent":false},"cited_by_percentile_year":null,"biblio":{"volume":null,"issue":null,"first_page":"4972","last_page":"4992"},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T13629","display_name":"Text Readability and Simplification","score":0.11219999939203262,"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/T13629","display_name":"Text Readability and Simplification","score":0.11219999939203262,"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/T10465","display_name":"Neurobiology of Language and Bilingualism","score":0.0803999975323677,"subfield":{"id":"https://openalex.org/subfields/2805","display_name":"Cognitive Neuroscience"},"field":{"id":"https://openalex.org/fields/28","display_name":"Neuroscience"},"domain":{"id":"https://openalex.org/domains/1","display_name":"Life Sciences"}},{"id":"https://openalex.org/T10730","display_name":"Language Development and Disorders","score":0.06270000338554382,"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/path","display_name":"Path (computing)","score":0.5597000122070312},{"id":"https://openalex.org/keywords/training","display_name":"Training (meteorology)","score":0.5449000000953674},{"id":"https://openalex.org/keywords/language-model","display_name":"Language model","score":0.2734000086784363},{"id":"https://openalex.org/keywords/action","display_name":"Action (physics)","score":0.2705000042915344},{"id":"https://openalex.org/keywords/training-set","display_name":"Training set","score":0.2619999945163727}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.5947999954223633},{"id":"https://openalex.org/C2777735758","wikidata":"https://www.wikidata.org/wiki/Q817765","display_name":"Path (computing)","level":2,"score":0.5597000122070312},{"id":"https://openalex.org/C2777211547","wikidata":"https://www.wikidata.org/wiki/Q17141490","display_name":"Training (meteorology)","level":2,"score":0.5449000000953674},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.4293000102043152},{"id":"https://openalex.org/C11413529","wikidata":"https://www.wikidata.org/wiki/Q8366","display_name":"Algorithm","level":1,"score":0.29989999532699585},{"id":"https://openalex.org/C137293760","wikidata":"https://www.wikidata.org/wiki/Q3621696","display_name":"Language model","level":2,"score":0.2734000086784363},{"id":"https://openalex.org/C2780791683","wikidata":"https://www.wikidata.org/wiki/Q846785","display_name":"Action (physics)","level":2,"score":0.2705000042915344},{"id":"https://openalex.org/C51632099","wikidata":"https://www.wikidata.org/wiki/Q3985153","display_name":"Training set","level":2,"score":0.2619999945163727},{"id":"https://openalex.org/C204321447","wikidata":"https://www.wikidata.org/wiki/Q30642","display_name":"Natural language processing","level":1,"score":0.24740000069141388},{"id":"https://openalex.org/C13662910","wikidata":"https://www.wikidata.org/wiki/Q193139","display_name":"Trajectory","level":2,"score":0.24289999902248383}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.18653/v1/2025.emnlp-main.251","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2025.emnlp-main.251","pdf_url":"https://aclanthology.org/2025.emnlp-main.251.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.251","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2025.emnlp-main.251","pdf_url":"https://aclanthology.org/2025.emnlp-main.251.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":[{"id":"https://openalex.org/G6854926366","display_name":null,"funder_award_id":"Tier 2","funder_id":"https://openalex.org/F4320320751","funder_display_name":"Ministry of Education - Singapore"},{"id":"https://openalex.org/G794643401","display_name":null,"funder_award_id":"MOE-T2EP20122-001","funder_id":"https://openalex.org/F4320320751","funder_display_name":"Ministry of Education - Singapore"}],"funders":[{"id":"https://openalex.org/F4320320671","display_name":"National Research Foundation","ror":"https://ror.org/05s0g1g46"},{"id":"https://openalex.org/F4320320709","display_name":"National Research Foundation Singapore","ror":"https://ror.org/03cpyc314"},{"id":"https://openalex.org/F4320320751","display_name":"Ministry of Education - Singapore","ror":"https://ror.org/01kcva023"},{"id":"https://openalex.org/F4320322724","display_name":"Ministry of Education, India","ror":"https://ror.org/048xjjh50"}],"has_content":{"grobid_xml":true,"pdf":true},"content_urls":{"pdf":"https://content.openalex.org/works/W4416036836.pdf","grobid_xml":"https://content.openalex.org/works/W4416036836.grobid-xml"},"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Long":[0,25,137],"chain-of-thought":[1],"(CoT)":[2],"supervision":[3],"has":[4],"become":[5],"a":[6,21],"common":[7,160],"strategy":[8],"to":[9,75,100,108,114],"enhance":[10],"reasoning":[11,180],"in":[12,28],"language":[13,31],"models.While":[14],"effective":[15,178],"for":[16,87,123,169,175],"large":[17],"models,":[18,91],"we":[19,23,54,83,134],"identify":[20],"phenomenon":[22],"call":[24],"CoT":[26,40,71,97,138,167],"Degradation,":[27],"which":[29],"small":[30,90],"models":[32,65],"(SLMs;":[33],"3B":[34],"parameters)":[35],"trained":[36,66],"on":[37,48,67,94],"limited":[38],"long":[39,70,96,166],"data":[41],"experience":[42],"significant":[43],"performance":[44,80,106],"deterioration.Through":[45],"extensive":[46],"experiments":[47],"the":[49,121,129,163],"Qwen2.5,LLaMA3":[50],"and":[51,171],"Gemma3":[52],"families,":[53],"demonstrate":[55],"that":[56,86,136],"this":[57,112,148],"degradation":[58],"is":[59],"widespread":[60],"across":[61],"SLMs.In":[62],"some":[63,88],"settings,":[64],"only":[68],"8k":[69],"examples":[72,98],"lose":[73],"up":[74],"75%":[76],"of":[77,131,165],"their":[78,104],"original":[79,105],"before":[81],"fine-tuning.Strikingly,":[82],"further":[84],"observe":[85],"particularly":[89],"even":[92],"training":[93,168],"220k":[95],"fails":[99],"recover":[101],"or":[102],"surpass":[103],"prior":[107],"fine-tuning.Our":[109],"analysis":[110],"attributes":[111],"effect":[113],"error":[115],"accumulation:":[116],"while":[117],"longer":[118],"responses":[119],"increase":[120],"capacity":[122],"multi-step":[124],"reasoning,":[125],"they":[126],"also":[127],"amplify":[128],"risk":[130],"compounding":[132],"mistakes.Furthermore,":[133],"find":[135],"Degradation":[139],"may":[140],"negatively":[141],"impacts":[142],"downstream":[143],"reinforcement":[144],"learning":[145],"(RL),":[146],"although":[147],"can":[149],"be":[150],"alleviated":[151],"by":[152],"sufficiently":[153],"scaled":[154],"supervised":[155],"fine-tuning":[156],"(SFT).Our":[157],"findings":[158],"challenge":[159],"assumptions":[161],"about":[162],"benefits":[164],"SLMs":[170],"offer":[172],"practical":[173],"guidance":[174],"building":[176],"more":[177],"small-scale":[179],"models.":[181]},"counts_by_year":[],"updated_date":"2026-04-15T08:11:43.952461","created_date":"2025-11-08T00:00:00"}