{"id":"https://openalex.org/W7118905480","doi":"https://doi.org/10.48550/arxiv.2601.02978","title":"Mechanistic Knobs in LLMs: Retrieving and Steering High-Order Semantic Features via Sparse Autoencoders","display_name":"Mechanistic Knobs in LLMs: Retrieving and Steering High-Order Semantic Features via Sparse Autoencoders","publication_year":2026,"publication_date":"2026-01-06","ids":{"openalex":"https://openalex.org/W7118905480","doi":"https://doi.org/10.48550/arxiv.2601.02978"},"language":null,"primary_location":{"id":"doi:10.48550/arxiv.2601.02978","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2601.02978","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,"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"},"type":"preprint","indexed_in":["datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://doi.org/10.48550/arxiv.2601.02978","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5121983989","display_name":"Ruikang Zhang","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Zhang, Ruikang","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5122009263","display_name":"Shuo Wang","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Wang, Shuo","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5030250226","display_name":"Q. Su","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Su, Qi","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]}],"institutions":[],"countries_distinct_count":0,"institutions_distinct_count":0,"corresponding_author_ids":[],"corresponding_institution_ids":[],"apc_list":null,"apc_paid":null,"fwci":null,"has_fulltext":false,"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":0.3431999981403351,"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":0.3431999981403351,"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":0.16120000183582306,"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":0.06759999692440033,"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":0.8057000041007996},{"id":"https://openalex.org/keywords/feature","display_name":"Feature (linguistics)","score":0.5069000124931335},{"id":"https://openalex.org/keywords/semantics","display_name":"Semantics (computer science)","score":0.46480000019073486},{"id":"https://openalex.org/keywords/semantic-feature","display_name":"Semantic feature","score":0.459199994802475},{"id":"https://openalex.org/keywords/pipeline","display_name":"Pipeline (software)","score":0.44830000400543213},{"id":"https://openalex.org/keywords/identification","display_name":"Identification (biology)","score":0.4253000020980835},{"id":"https://openalex.org/keywords/control","display_name":"Control (management)","score":0.3653999865055084},{"id":"https://openalex.org/keywords/stability","display_name":"Stability (learning theory)","score":0.36230000853538513}],"concepts":[{"id":"https://openalex.org/C2781067378","wikidata":"https://www.wikidata.org/wiki/Q17027399","display_name":"Interpretability","level":2,"score":0.8057000041007996},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.7468000054359436},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.5746999979019165},{"id":"https://openalex.org/C2776401178","wikidata":"https://www.wikidata.org/wiki/Q12050496","display_name":"Feature (linguistics)","level":2,"score":0.5069000124931335},{"id":"https://openalex.org/C204321447","wikidata":"https://www.wikidata.org/wiki/Q30642","display_name":"Natural language processing","level":1,"score":0.5041999816894531},{"id":"https://openalex.org/C184337299","wikidata":"https://www.wikidata.org/wiki/Q1437428","display_name":"Semantics (computer science)","level":2,"score":0.46480000019073486},{"id":"https://openalex.org/C2781122975","wikidata":"https://www.wikidata.org/wiki/Q16928266","display_name":"Semantic feature","level":2,"score":0.459199994802475},{"id":"https://openalex.org/C43521106","wikidata":"https://www.wikidata.org/wiki/Q2165493","display_name":"Pipeline (software)","level":2,"score":0.44830000400543213},{"id":"https://openalex.org/C116834253","wikidata":"https://www.wikidata.org/wiki/Q2039217","display_name":"Identification (biology)","level":2,"score":0.4253000020980835},{"id":"https://openalex.org/C2775924081","wikidata":"https://www.wikidata.org/wiki/Q55608371","display_name":"Control (management)","level":2,"score":0.3653999865055084},{"id":"https://openalex.org/C112972136","wikidata":"https://www.wikidata.org/wiki/Q7595718","display_name":"Stability (learning theory)","level":2,"score":0.36230000853538513},{"id":"https://openalex.org/C61797465","wikidata":"https://www.wikidata.org/wiki/Q1188986","display_name":"Term (time)","level":2,"score":0.3434999883174896},{"id":"https://openalex.org/C137293760","wikidata":"https://www.wikidata.org/wiki/Q3621696","display_name":"Language model","level":2,"score":0.3377000093460083},{"id":"https://openalex.org/C114289077","wikidata":"https://www.wikidata.org/wiki/Q3284399","display_name":"Statistical model","level":2,"score":0.3098999857902527},{"id":"https://openalex.org/C52622490","wikidata":"https://www.wikidata.org/wiki/Q1026626","display_name":"Feature extraction","level":2,"score":0.30149999260902405},{"id":"https://openalex.org/C153180895","wikidata":"https://www.wikidata.org/wiki/Q7148389","display_name":"Pattern recognition (psychology)","level":2,"score":0.2987000048160553},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.2750999927520752},{"id":"https://openalex.org/C2780801425","wikidata":"https://www.wikidata.org/wiki/Q5164392","display_name":"Construct (python library)","level":2,"score":0.27090001106262207},{"id":"https://openalex.org/C2777735758","wikidata":"https://www.wikidata.org/wiki/Q817765","display_name":"Path (computing)","level":2,"score":0.2669000029563904},{"id":"https://openalex.org/C195324797","wikidata":"https://www.wikidata.org/wiki/Q33742","display_name":"Natural language","level":2,"score":0.2554999887943268},{"id":"https://openalex.org/C89611455","wikidata":"https://www.wikidata.org/wiki/Q6804646","display_name":"Mechanism (biology)","level":2,"score":0.25369998812675476},{"id":"https://openalex.org/C197914299","wikidata":"https://www.wikidata.org/wiki/Q18650","display_name":"Semantic memory","level":3,"score":0.2500999867916107}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.48550/arxiv.2601.02978","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2601.02978","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,"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":null,"raw_source_name":null,"raw_type":"Preprint"}],"best_oa_location":{"id":"doi:10.48550/arxiv.2601.02978","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2601.02978","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,"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"},"sustainable_development_goals":[],"awards":[],"funders":[],"has_content":{"grobid_xml":false,"pdf":false},"content_urls":null,"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Recent":[0],"work":[1],"in":[2,15,25,39],"Mechanistic":[3],"Interpretability":[4],"(MI)":[5],"has":[6],"enabled":[7],"the":[8,31,94,161,185],"identification":[9],"and":[10,53,81,119,152,177],"intervention":[11],"of":[12,34,112,174,187],"internal":[13,28,57,148],"features":[14,29,58,88],"Large":[16],"Language":[17],"Models":[18],"(LLMs).":[19],"However,":[20],"a":[21,47,67,100,146,179],"persistent":[22],"challenge":[23],"lies":[24],"linking":[26],"such":[27],"to":[30,84,122],"reliable":[32],"control":[33],"complex,":[35],"behavior-level":[36],"semantic":[37,75,163],"attributes":[38],"language":[40],"generation.":[41],"In":[42],"this":[43],"paper,":[44],"we":[45,103,139],"propose":[46],"Sparse":[48],"Autoencoder-based":[49],"framework":[50],"for":[51,184],"retrieving":[52],"steering":[54,111,125],"semantically":[55],"interpretable":[56],"associated":[59],"with":[60,160],"high-level":[61],"linguistic":[62,157],"behaviors.":[63,190],"Our":[64,165],"method":[65,107],"employs":[66],"contrastive":[68],"feature":[69,149],"retrieval":[70],"pipeline":[71],"based":[72],"on":[73,145],"controlled":[74],"oppositions,":[76],"combing":[77],"statistical":[78],"activation":[79,91,124],"analysis":[80],"generation-based":[82],"validation":[83],"distill":[85],"monosemantic":[86],"functional":[87],"from":[89],"sparse":[90],"spaces.":[92],"Using":[93],"Big":[95],"Five":[96],"personality":[97],"traits":[98],"as":[99],"case":[101],"study,":[102],"demonstrate":[104],"that":[105,168],"our":[106],"enables":[108],"precise,":[109],"bidirectional":[110],"model":[113],"behavior":[114],"while":[115],"maintaining":[116],"superior":[117],"stability":[118],"performance":[120],"compared":[121],"existing":[123],"methods":[126],"like":[127],"Contrastive":[128],"Activation":[129],"Addition":[130],"(CAA).":[131],"We":[132],"further":[133],"identify":[134],"an":[135],"empirical":[136],"effect,":[137],"which":[138],"term":[140],"Functional":[141],"Faithfulness,":[142],"whereby":[143],"intervening":[144],"specific":[147],"induces":[150],"coherent":[151],"predictable":[153],"shifts":[154],"across":[155],"multiple":[156],"dimensions":[158],"aligned":[159],"target":[162],"attribute.":[164],"findings":[166],"suggest":[167],"LLMs":[169],"internalize":[170],"deeply":[171],"integrated":[172],"representations":[173],"high-order":[175],"concepts,":[176],"provide":[178],"novel,":[180],"robust":[181],"mechanistic":[182],"path":[183],"regulation":[186],"complex":[188],"AI":[189]},"counts_by_year":[],"updated_date":"2026-07-28T07:46:37.118299","created_date":"2026-01-08T00:00:00"}