{"id":"https://openalex.org/W7138029733","doi":"https://doi.org/10.1609/aaai.v40i36.40281","title":"How Does Chain of Thought Think? Mechanistic Interpretability of Chain-of-Thought Reasoning with Sparse Autoencoding","display_name":"How Does Chain of Thought Think? Mechanistic Interpretability of Chain-of-Thought Reasoning with Sparse Autoencoding","publication_year":2026,"publication_date":"2026-03-14","ids":{"openalex":"https://openalex.org/W7138029733","doi":"https://doi.org/10.1609/aaai.v40i36.40281"},"language":null,"primary_location":{"id":"doi:10.1609/aaai.v40i36.40281","is_oa":true,"landing_page_url":"https://doi.org/10.1609/aaai.v40i36.40281","pdf_url":null,"source":{"id":"https://openalex.org/S4210191458","display_name":"Proceedings of the AAAI Conference on Artificial Intelligence","issn_l":"2159-5399","issn":["2159-5399","2374-3468"],"is_oa":true,"is_in_doaj":false,"is_core":false,"host_organization":"https://openalex.org/P4310320058","host_organization_name":"Association for the Advancement of Artificial Intelligence","host_organization_lineage":["https://openalex.org/P4310320058"],"host_organization_lineage_names":["Association for the Advancement of Artificial Intelligence"],"type":"conference"},"license":null,"license_id":null,"version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Proceedings of the AAAI Conference on Artificial Intelligence","raw_type":"journal-article"},"type":"conference-paper","indexed_in":["crossref"],"open_access":{"is_oa":true,"oa_status":"diamond","oa_url":"https://doi.org/10.1609/aaai.v40i36.40281","any_repository_has_fulltext":false},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5129687665","display_name":"Xi Chen","orcid":null},"institutions":[{"id":"https://openalex.org/I121797337","display_name":"Leiden University","ror":"https://ror.org/027bh9e22","country_code":"NL","type":"education","lineage":["https://openalex.org/I121797337"]}],"countries":["NL"],"is_corresponding":false,"raw_author_name":"Xi Chen","raw_affiliation_strings":["Leiden University"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Leiden University","institution_ids":["https://openalex.org/I121797337"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5129670985","display_name":"Aske Plaat","orcid":null},"institutions":[{"id":"https://openalex.org/I121797337","display_name":"Leiden University","ror":"https://ror.org/027bh9e22","country_code":"NL","type":"education","lineage":["https://openalex.org/I121797337"]}],"countries":["NL"],"is_corresponding":false,"raw_author_name":"Aske Plaat","raw_affiliation_strings":["Leiden University"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Leiden University","institution_ids":["https://openalex.org/I121797337"]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5129652815","display_name":"Niki van Stein","orcid":null},"institutions":[{"id":"https://openalex.org/I121797337","display_name":"Leiden University","ror":"https://ror.org/027bh9e22","country_code":"NL","type":"education","lineage":["https://openalex.org/I121797337"]}],"countries":["NL"],"is_corresponding":false,"raw_author_name":"Niki Van Stein","raw_affiliation_strings":["Leiden University"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Leiden University","institution_ids":["https://openalex.org/I121797337"]}]}],"institutions":[],"countries_distinct_count":1,"institutions_distinct_count":1,"corresponding_author_ids":[],"corresponding_institution_ids":["https://openalex.org/I121797337"],"apc_list":null,"apc_paid":null,"fwci":0.0,"has_fulltext":false,"cited_by_count":0,"citation_normalized_percentile":{"value":0.19239374,"is_in_top_1_percent":false,"is_in_top_10_percent":false},"cited_by_percentile_year":null,"biblio":{"volume":"40","issue":"36","first_page":"30297","last_page":"30305"},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T10465","display_name":"Neurobiology of Language and Bilingualism","score":0.23319999873638153,"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"}},"topics":[{"id":"https://openalex.org/T10465","display_name":"Neurobiology of Language and Bilingualism","score":0.23319999873638153,"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/T10028","display_name":"Topic Modeling","score":0.21899999678134918,"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/T11714","display_name":"Multimodal Machine Learning Applications","score":0.03229999914765358,"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"}}],"keywords":[{"id":"https://openalex.org/keywords/interpretability","display_name":"Interpretability","score":0.9668999910354614},{"id":"https://openalex.org/keywords/feature","display_name":"Feature (linguistics)","score":0.5878000259399414},{"id":"https://openalex.org/keywords/set","display_name":"Set (abstract data type)","score":0.5659000277519226},{"id":"https://openalex.org/keywords/modular-design","display_name":"Modular design","score":0.5479999780654907},{"id":"https://openalex.org/keywords/causal-chain","display_name":"Causal chain","score":0.4648999869823456},{"id":"https://openalex.org/keywords/process","display_name":"Process (computing)","score":0.44929999113082886},{"id":"https://openalex.org/keywords/contrast","display_name":"Contrast (vision)","score":0.44279998540878296}],"concepts":[{"id":"https://openalex.org/C2781067378","wikidata":"https://www.wikidata.org/wiki/Q17027399","display_name":"Interpretability","level":2,"score":0.9668999910354614},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.6718000173568726},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.6324999928474426},{"id":"https://openalex.org/C2776401178","wikidata":"https://www.wikidata.org/wiki/Q12050496","display_name":"Feature (linguistics)","level":2,"score":0.5878000259399414},{"id":"https://openalex.org/C177264268","wikidata":"https://www.wikidata.org/wiki/Q1514741","display_name":"Set (abstract data type)","level":2,"score":0.5659000277519226},{"id":"https://openalex.org/C101468663","wikidata":"https://www.wikidata.org/wiki/Q1620158","display_name":"Modular design","level":2,"score":0.5479999780654907},{"id":"https://openalex.org/C79897977","wikidata":"https://www.wikidata.org/wiki/Q5054568","display_name":"Causal chain","level":2,"score":0.4648999869823456},{"id":"https://openalex.org/C98045186","wikidata":"https://www.wikidata.org/wiki/Q205663","display_name":"Process (computing)","level":2,"score":0.44929999113082886},{"id":"https://openalex.org/C2776502983","wikidata":"https://www.wikidata.org/wiki/Q690182","display_name":"Contrast (vision)","level":2,"score":0.44279998540878296},{"id":"https://openalex.org/C2778067643","wikidata":"https://www.wikidata.org/wiki/Q166507","display_name":"Interval (graph theory)","level":2,"score":0.4327000081539154},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.4092000126838684},{"id":"https://openalex.org/C204321447","wikidata":"https://www.wikidata.org/wiki/Q30642","display_name":"Natural language processing","level":1,"score":0.3767000138759613},{"id":"https://openalex.org/C153180895","wikidata":"https://www.wikidata.org/wiki/Q7148389","display_name":"Pattern recognition (psychology)","level":2,"score":0.3382999897003174},{"id":"https://openalex.org/C199185054","wikidata":"https://www.wikidata.org/wiki/Q552299","display_name":"Chain (unit)","level":2,"score":0.33390000462532043},{"id":"https://openalex.org/C2776639384","wikidata":"https://www.wikidata.org/wiki/Q840396","display_name":"Ideal (ethics)","level":2,"score":0.31150001287460327},{"id":"https://openalex.org/C11671645","wikidata":"https://www.wikidata.org/wiki/Q5054567","display_name":"Causal model","level":2,"score":0.26460000872612},{"id":"https://openalex.org/C19768560","wikidata":"https://www.wikidata.org/wiki/Q320727","display_name":"Dependency (UML)","level":2,"score":0.25679999589920044}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.1609/aaai.v40i36.40281","is_oa":true,"landing_page_url":"https://doi.org/10.1609/aaai.v40i36.40281","pdf_url":null,"source":{"id":"https://openalex.org/S4210191458","display_name":"Proceedings of the AAAI Conference on Artificial Intelligence","issn_l":"2159-5399","issn":["2159-5399","2374-3468"],"is_oa":true,"is_in_doaj":false,"is_core":false,"host_organization":"https://openalex.org/P4310320058","host_organization_name":"Association for the Advancement of Artificial Intelligence","host_organization_lineage":["https://openalex.org/P4310320058"],"host_organization_lineage_names":["Association for the Advancement of Artificial Intelligence"],"type":"conference"},"license":null,"license_id":null,"version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Proceedings of the AAAI Conference on Artificial Intelligence","raw_type":"journal-article"}],"best_oa_location":{"id":"doi:10.1609/aaai.v40i36.40281","is_oa":true,"landing_page_url":"https://doi.org/10.1609/aaai.v40i36.40281","pdf_url":null,"source":{"id":"https://openalex.org/S4210191458","display_name":"Proceedings of the AAAI Conference on Artificial Intelligence","issn_l":"2159-5399","issn":["2159-5399","2374-3468"],"is_oa":true,"is_in_doaj":false,"is_core":false,"host_organization":"https://openalex.org/P4310320058","host_organization_name":"Association for the Advancement of Artificial Intelligence","host_organization_lineage":["https://openalex.org/P4310320058"],"host_organization_lineage_names":["Association for the Advancement of Artificial Intelligence"],"type":"conference"},"license":null,"license_id":null,"version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Proceedings of the AAAI Conference on Artificial Intelligence","raw_type":"journal-article"},"sustainable_development_goals":[{"id":"https://metadata.un.org/sdg/4","display_name":"Quality Education","score":0.5394716858863831}],"awards":[],"funders":[],"has_content":{"grobid_xml":false,"pdf":false},"content_urls":null,"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Chain\u2011of\u2011thought":[0],"(CoT)":[1],"prompting":[2,173],"boosts":[3],"Large":[4],"Language":[5],"Models":[6],"accuracy":[7],"on":[8],"multi\u2011step":[9],"tasks,":[10],"yet":[11],"whether":[12],"the":[13,17,26,76,107,117,146],"generated":[14],"``thoughts''":[15],"reflect":[16],"true":[18],"internal":[19,113,162],"reasoning":[20],"process":[21],"is":[22,141],"unresolved.":[23],"We":[24,129],"present":[25,144],"first":[27],"feature\u2011level":[28],"causal":[29],"study":[30],"of":[31,64],"CoT":[32,55,94,139,157],"faithfulness.":[33],"Combining":[34],"sparse":[35],"autoencoders":[36],"with":[37],"activation":[38,100],"patching,":[39],"we":[40],"extract":[41],"monosemantic":[42],"features":[43,66],"from":[44,125],"Pythia\u201170M":[45],"and":[46,56,102,132],"Pythia\u20112.8B":[47],"while":[48],"they":[49],"tackle":[50],"GSM8K":[51],"math":[52],"problems":[53],"under":[54],"plain":[57],"(noCoT)":[58],"prompting.":[59],"Swapping":[60],"a":[61,68,87,171],"small":[62],"set":[63],"CoT\u2011reasoning":[65],"into":[67],"noCoT":[69],"run":[70],"raises":[71],"answer":[72],"log\u2011probabilities":[73],"significantly":[74,98],"in":[75,84,106,120,145,164],"2.8B":[77],"model,":[78,109],"but":[79,149],"has":[80],"no":[81],"reliable":[82],"effect":[83],"70M,":[85],"revealing":[86],"clear":[88],"contrast":[89],"for":[90],"these":[91],"two":[92],"scales.":[93],"also":[95],"leads":[96],"to":[97,127],"higher":[99],"sparsity":[101],"feature":[103],"interpretability":[104],"scores":[105],"larger":[108],"signalling":[110],"more":[111,160],"modular":[112],"computation.":[114],"For":[115],"example,":[116],"model's":[118],"confidence":[119],"generating":[121],"correct":[122],"answers":[123],"improves":[124],"1.2":[126],"4.3.":[128],"introduce":[130],"patch\u2011curves":[131],"random\u2011feature":[133],"patching":[134],"baselines,":[135],"showing":[136],"that":[137,156],"useful":[138],"information":[140],"not":[142],"only":[143],"top-K":[147],"patches":[148],"widely":[150],"distributed.":[151],"Overall,":[152],"our":[153],"results":[154],"indicate":[155],"can":[158],"induce":[159],"interpretable":[161],"structures":[163],"high-capacity":[165],"LLMs,":[166],"validating":[167],"its":[168],"role":[169],"as":[170],"structured":[172],"method.":[174]},"counts_by_year":[],"updated_date":"2026-07-29T14:22:42.915294","created_date":"2026-03-18T00:00:00"}