{"id":"https://openalex.org/W7202210108","doi":"https://doi.org/10.48550/arxiv.2608.08168","title":"Thinking vs. NoThinking: Towards Interpreting Reasoning Mechanisms of Large Language Models via Sparse Autoencoders","display_name":"Thinking vs. NoThinking: Towards Interpreting Reasoning Mechanisms of Large Language Models via Sparse Autoencoders","publication_year":2026,"publication_date":"2026-08-08","ids":{"openalex":"https://openalex.org/W7202210108","doi":"https://doi.org/10.48550/arxiv.2608.08168"},"language":"en","primary_location":{"id":"pmh:oai:arXiv.org:2608.08168","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2608.08168","pdf_url":"https://arxiv.org/pdf/2608.08168","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":null,"license_id":null,"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/2608.08168","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5125727944","display_name":"Bo Cheng","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I194450716","display_name":"Jilin University","ror":"https://ror.org/00js3aw79","country_code":"CN","type":"education","lineage":["https://openalex.org/I194450716"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Cheng, Bo","raw_affiliation_strings":["School of Artificial Intelligence , Jilin University"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"School of Artificial Intelligence , Jilin University","institution_ids":["https://openalex.org/I194450716"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5146443720","display_name":"Qiaolin Lu","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I14243506","display_name":"Hong Kong Polytechnic University","ror":"https://ror.org/0030zas98","country_code":"HK","type":"education","lineage":["https://openalex.org/I14243506"]}],"countries":["HK"],"is_corresponding":false,"raw_author_name":"Lu, Qiaolin","raw_affiliation_strings":["The Hong Kong Polytechnic University"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"The Hong Kong Polytechnic University","institution_ids":["https://openalex.org/I14243506"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5146454025","display_name":"Yi Chang","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Chang, Yi","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5137917501","display_name":"Yuan Wu","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I194450716","display_name":"Jilin University","ror":"https://ror.org/00js3aw79","country_code":"CN","type":"education","lineage":["https://openalex.org/I194450716"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Wu, Yuan","raw_affiliation_strings":["School of Artificial Intelligence , Jilin University"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"School of Artificial Intelligence , Jilin University","institution_ids":["https://openalex.org/I194450716"]}]}],"institutions":[],"countries_distinct_count":2,"institutions_distinct_count":2,"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/T12026","display_name":"Explainable Artificial Intelligence (XAI)","score":0.6215999722480774,"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/T12026","display_name":"Explainable Artificial Intelligence (XAI)","score":0.6215999722480774,"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/T10028","display_name":"Topic Modeling","score":0.3400999903678894,"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":0.008200000040233135,"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":0.902999997138977},{"id":"https://openalex.org/keywords/chain-of-thought-reasoning","display_name":"chain-of-thought reasoning","score":0.9599999785423279},{"id":"https://openalex.org/keywords/sparse-autoencoders","display_name":"sparse autoencoders","score":0.9470000267028809},{"id":"https://openalex.org/keywords/model-interpretability","display_name":"model interpretability","score":0.5559999942779541},{"id":"https://openalex.org/keywords/mathematical-problem-solving","display_name":"mathematical problem solving","score":0.6010000109672546},{"id":"https://openalex.org/keywords/metacognition","display_name":"metacognition","score":0.5929999947547913},{"id":"https://openalex.org/keywords/overgeneration","display_name":"overgeneration","score":0.6980000138282776},{"id":"https://openalex.org/keywords/symbolic-computation","display_name":"symbolic computation","score":0.6079999804496765},{"id":"https://openalex.org/keywords/syntactic-structure","display_name":"syntactic structure","score":0.765999972820282}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.5870000123977661},{"id":"https://openalex.org/C169900460","wikidata":"https://www.wikidata.org/wiki/Q2200417","display_name":"Cognition","level":2,"score":0.536899983882904},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.5026000142097473},{"id":"https://openalex.org/C2776401178","wikidata":"https://www.wikidata.org/wiki/Q12050496","display_name":"Feature (linguistics)","level":2,"score":0.4438999891281128},{"id":"https://openalex.org/C180747234","wikidata":"https://www.wikidata.org/wiki/Q23373","display_name":"Cognitive psychology","level":1,"score":0.41530001163482666},{"id":"https://openalex.org/C204321447","wikidata":"https://www.wikidata.org/wiki/Q30642","display_name":"Natural language processing","level":1,"score":0.41359999775886536},{"id":"https://openalex.org/C188147891","wikidata":"https://www.wikidata.org/wiki/Q147638","display_name":"Cognitive science","level":1,"score":0.4004000127315521},{"id":"https://openalex.org/C48677424","wikidata":"https://www.wikidata.org/wiki/Q6888088","display_name":"Mode (computer interface)","level":2,"score":0.37209999561309814},{"id":"https://openalex.org/C183521366","wikidata":"https://www.wikidata.org/wiki/Q7256422","display_name":"Psychology of reasoning","level":4,"score":0.35910001397132874},{"id":"https://openalex.org/C15744967","wikidata":"https://www.wikidata.org/wiki/Q9418","display_name":"Psychology","level":0,"score":0.3163999915122986},{"id":"https://openalex.org/C36964233","wikidata":"https://www.wikidata.org/wiki/Q7920942","display_name":"Verbal reasoning","level":3,"score":0.3073999881744385},{"id":"https://openalex.org/C12713177","wikidata":"https://www.wikidata.org/wiki/Q1900281","display_name":"Perspective (graphical)","level":2,"score":0.3018999993801117},{"id":"https://openalex.org/C137293760","wikidata":"https://www.wikidata.org/wiki/Q3621696","display_name":"Language model","level":2,"score":0.2994000017642975},{"id":"https://openalex.org/C2983448237","wikidata":"https://www.wikidata.org/wiki/Q1078276","display_name":"Language understanding","level":2,"score":0.27489998936653137},{"id":"https://openalex.org/C89267518","wikidata":"https://www.wikidata.org/wiki/Q179488","display_name":"Psycholinguistics","level":3,"score":0.2646999955177307}],"mesh":[],"locations_count":2,"locations":[{"id":"pmh:oai:arXiv.org:2608.08168","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2608.08168","pdf_url":"https://arxiv.org/pdf/2608.08168","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":null,"license_id":null,"version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},{"id":"doi:10.48550/arxiv.2608.08168","doi":"https://doi.org/10.48550/arxiv.2608.08168","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2608.08168","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":null,"license_id":null,"version":null,"is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"Preprint"}],"best_oa_location":{"id":"pmh:oai:arXiv.org:2608.08168","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2608.08168","pdf_url":"https://arxiv.org/pdf/2608.08168","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":null,"license_id":null,"version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},"sustainable_development_goals":[],"study_designs":[],"awards":[],"funders":[],"has_content":{"pdf":true,"grobid_xml":true},"content_urls":{"pdf":"https://content.openalex.org/works/W7202210108.pdf","grobid_xml":"https://content.openalex.org/works/W7202210108.grobid-xml"},"referenced_works_count":0,"referenced_works":[],"related_works":["https://openalex.org/W6966980334","https://openalex.org/W7124175571","https://openalex.org/W7102624217","https://openalex.org/W4414930565","https://openalex.org/W7125673123","https://openalex.org/W4416887915","https://openalex.org/W7133530652","https://openalex.org/W7166333208","https://openalex.org/W7204942450","https://openalex.org/W4415328648"],"abstract_inverted_index":{"While":[0],"Large":[1],"Language":[2],"Models":[3],"(LLMs)":[4],"employing":[5],"Chain-of-Thought":[6],"(CoT)":[7],"exhibit":[8],"superior":[9],"reasoning":[10,73,119],"capabilities,":[11],"the":[12,41,48,68,105],"neural":[13],"mechanisms":[14],"distinguishing":[15],"this":[16,31],"explicit":[17],"Thinking":[18,75,135],"mode":[19,76,93],"from":[20],"direct":[21],"answer":[22],"generation":[23],"(NoThinking":[24],"mode)":[25],"remain":[26],"poorly":[27],"understood.":[28],"To":[29],"deconstruct":[30],"cognitive":[32],"process,":[33],"we":[34,61],"apply":[35],"Top-K":[36],"Sparse":[37],"Autoencoders":[38],"(SAEs)":[39],"to":[40,137],"intermediate":[42],"representations":[43],"of":[44,55,88],"DeepSeek-R1-Distill-Qwen-7B":[45],"and":[46,80,97,120,131,147,151],"examine":[47],"model's":[49],"divergent":[50],"behaviors":[51],"across":[52],"math-solving":[53],"tasks":[54],"three":[56,106,116],"distinct":[57,165],"difficulty":[58],"levels.":[59],"Observationally,":[60],"identify":[62],"a":[63,158,171],"clear":[64],"distinction":[65],"in":[66],"how":[67],"model":[69],"functions":[70],"under":[71,168],"two":[72],"modes:":[74],"relies":[77],"on":[78,157],"sparse":[79,109],"high-intensity":[81],"feature":[82],"activations":[83],"driving":[84],"verbal":[85],"deduction":[86],"independent":[87],"problem":[89],"complexity,":[90],"whereas":[91],"NoThinking":[92],"exhibits":[94],"an":[95],"adaptive":[96],"diffuse":[98],"pattern":[99],"prioritizing":[100],"symbolic":[101],"manipulation.":[102],"Causally,":[103],"suppressing":[104],"most":[107],"active":[108],"features":[110],"by":[111,143],"Total":[112],"Activation":[113],"Volume":[114],"reveals":[115],"principles:":[117],"(i)":[118],"syntactic":[121],"structure":[122],"are":[123],"tightly":[124],"coupled,":[125],"as":[126],"interventions":[127],"consistently":[128,172],"degrade":[129],"\\LaTeX{}":[130],"boxed-solution":[132],"formatting;":[133],"(ii)":[134],"responds":[136],"disruption":[138],"with":[139],"compensatory":[140],"over-generation":[141],"marked":[142],"increased":[144],"metacognitive":[145],"cues":[146],"repetitive,":[148],"low-information":[149],"continuations;":[150],"(iii)":[152],"coherent":[153],"CoT":[154],"behavior":[155],"depends":[156],"fragile":[159],"coordination":[160],"among":[161],"specialized":[162],"features,":[163],"yielding":[164],"failure":[166],"modes":[167],"perturbation":[169],"but":[170],"impaired":[173],"output":[174],"structure.":[175]},"counts_by_year":[],"updated_date":"2026-10-28T10:33:24.094632","created_date":"2026-08-12T00:00:00"}