{"id":"https://openalex.org/W7164913829","doi":"https://doi.org/10.48550/arxiv.2606.16920","title":"Demystifying Variance in Circuit Discovery of LLMs","display_name":"Demystifying Variance in Circuit Discovery of LLMs","publication_year":2026,"publication_date":"2026-06-15","ids":{"openalex":"https://openalex.org/W7164913829","doi":"https://doi.org/10.48550/arxiv.2606.16920"},"language":null,"primary_location":{"id":"doi:10.48550/arxiv.2606.16920","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2606.16920","pdf_url":null,"source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":null,"issn":null,"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.2606.16920","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5101542416","display_name":"F. Wu","orcid":"https://orcid.org/0000-0003-2739-2377"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Wu, Frank Zhengqing","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5138754746","display_name":"Francesco Tonin","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Tonin, Francesco","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5138740439","display_name":"Volkan Cevher","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Cevher, Volkan","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":0.5185999870300293,"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":0.5185999870300293,"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/T11948","display_name":"Machine Learning in Materials Science","score":0.19290000200271606,"subfield":{"id":"https://openalex.org/subfields/2505","display_name":"Materials Chemistry"},"field":{"id":"https://openalex.org/fields/25","display_name":"Materials Science"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}},{"id":"https://openalex.org/T11273","display_name":"Advanced Graph Neural Networks","score":0.04749999940395355,"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/resampling","display_name":"Resampling","score":0.6992999911308289},{"id":"https://openalex.org/keywords/interpretability","display_name":"Interpretability","score":0.6195999979972839},{"id":"https://openalex.org/keywords/variance","display_name":"Variance (accounting)","score":0.6108999848365784},{"id":"https://openalex.org/keywords/metric","display_name":"Metric (unit)","score":0.5048999786376953},{"id":"https://openalex.org/keywords/population","display_name":"Population","score":0.5033000111579895},{"id":"https://openalex.org/keywords/similarity","display_name":"Similarity (geometry)","score":0.44110000133514404},{"id":"https://openalex.org/keywords/task","display_name":"Task (project management)","score":0.3328000009059906},{"id":"https://openalex.org/keywords/electronic-circuit","display_name":"Electronic circuit","score":0.3246999979019165}],"concepts":[{"id":"https://openalex.org/C150921843","wikidata":"https://www.wikidata.org/wiki/Q1170431","display_name":"Resampling","level":2,"score":0.6992999911308289},{"id":"https://openalex.org/C2781067378","wikidata":"https://www.wikidata.org/wiki/Q17027399","display_name":"Interpretability","level":2,"score":0.6195999979972839},{"id":"https://openalex.org/C196083921","wikidata":"https://www.wikidata.org/wiki/Q7915758","display_name":"Variance (accounting)","level":2,"score":0.6108999848365784},{"id":"https://openalex.org/C176217482","wikidata":"https://www.wikidata.org/wiki/Q860554","display_name":"Metric (unit)","level":2,"score":0.5048999786376953},{"id":"https://openalex.org/C2908647359","wikidata":"https://www.wikidata.org/wiki/Q2625603","display_name":"Population","level":2,"score":0.5033000111579895},{"id":"https://openalex.org/C103278499","wikidata":"https://www.wikidata.org/wiki/Q254465","display_name":"Similarity (geometry)","level":3,"score":0.44110000133514404},{"id":"https://openalex.org/C105795698","wikidata":"https://www.wikidata.org/wiki/Q12483","display_name":"Statistics","level":1,"score":0.3806999921798706},{"id":"https://openalex.org/C149782125","wikidata":"https://www.wikidata.org/wiki/Q160039","display_name":"Econometrics","level":1,"score":0.3653999865055084},{"id":"https://openalex.org/C33923547","wikidata":"https://www.wikidata.org/wiki/Q395","display_name":"Mathematics","level":0,"score":0.358599990606308},{"id":"https://openalex.org/C2780451532","wikidata":"https://www.wikidata.org/wiki/Q759676","display_name":"Task (project management)","level":2,"score":0.3328000009059906},{"id":"https://openalex.org/C134146338","wikidata":"https://www.wikidata.org/wiki/Q1815901","display_name":"Electronic circuit","level":2,"score":0.3246999979019165},{"id":"https://openalex.org/C99476002","wikidata":"https://www.wikidata.org/wiki/Q42297","display_name":"Analysis of variance","level":2,"score":0.32429999113082886},{"id":"https://openalex.org/C26517878","wikidata":"https://www.wikidata.org/wiki/Q228039","display_name":"Key (lock)","level":2,"score":0.30320000648498535},{"id":"https://openalex.org/C183115368","wikidata":"https://www.wikidata.org/wiki/Q856577","display_name":"Weighting","level":2,"score":0.3009999990463257},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.29649999737739563},{"id":"https://openalex.org/C2780009758","wikidata":"https://www.wikidata.org/wiki/Q6804172","display_name":"Measure (data warehouse)","level":2,"score":0.28299999237060547},{"id":"https://openalex.org/C21308566","wikidata":"https://www.wikidata.org/wiki/Q7169365","display_name":"Permutation (music)","level":2,"score":0.2678999900817871},{"id":"https://openalex.org/C48103436","wikidata":"https://www.wikidata.org/wiki/Q599031","display_name":"State (computer science)","level":2,"score":0.2630000114440918},{"id":"https://openalex.org/C89611455","wikidata":"https://www.wikidata.org/wiki/Q6804646","display_name":"Mechanism (biology)","level":2,"score":0.26030001044273376},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.2572999894618988},{"id":"https://openalex.org/C153180895","wikidata":"https://www.wikidata.org/wiki/Q7148389","display_name":"Pattern recognition (psychology)","level":2,"score":0.2572000026702881},{"id":"https://openalex.org/C124101348","wikidata":"https://www.wikidata.org/wiki/Q172491","display_name":"Data mining","level":1,"score":0.2556000053882599},{"id":"https://openalex.org/C193244246","wikidata":"https://www.wikidata.org/wiki/Q5432696","display_name":"False discovery rate","level":3,"score":0.2547999918460846}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.48550/arxiv.2606.16920","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2606.16920","pdf_url":null,"source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":null,"issn":null,"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.2606.16920","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2606.16920","pdf_url":null,"source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":null,"issn":null,"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":[{"display_name":"No poverty","score":0.7197579741477966,"id":"https://metadata.un.org/sdg/1"}],"awards":[],"funders":[],"has_content":{"grobid_xml":false,"pdf":false},"content_urls":null,"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Circuit":[0],"discovery":[1,106],"is":[2,210,222,239],"a":[3,19,52,77,113,153,164,245],"key":[4],"technique":[5],"in":[6,86,138,170,228],"mechanistic":[7],"interpretability":[8],"to":[9,134,144,151,180,190,199],"pinpoint":[10],"the":[11,23,31,45,58,64,69,94,139,160,229,235,251],"model":[12],"components":[13],"that":[14,101,108,124,146,156,174,184,208,234,248],"are":[15,71],"crucial":[16],"for":[17,250],"performing":[18],"given":[20],"task.":[21],"Although":[22],"current":[24],"state-of-the-art":[25],"method":[26,107],"(EAP-IG)":[27],"performs":[28],"well":[29],"on":[30,163],"metric":[32],"of":[33,55,96,237],"(un)faithfulness,":[34],"it":[35,147,209],"suffers":[36],"from":[37,57,219,226],"substantial":[38],"variability.":[39],"This":[40,91,141],"includes":[41],"resampling":[42,119],"variance,":[43,62,75,205],"where":[44,63,76],"circuit":[46,66,78,105,155],"changes":[47],"when":[48,68],"we":[49,206],"probe":[50],"with":[51,79,112,130],"new":[53,104],"batch":[54],"data":[56],"same":[59],"distribution;":[60],"rephrasing":[61,125],"discovered":[65],"shifts":[67],"prompts":[70,129],"rephrased;":[72],"and":[73,158,194],"sample-wise":[74,204],"low":[80],"population":[81],"unfaithfulness":[82,87,215,221,238],"exhibits":[83],"large":[84],"fluctuations":[85],"across":[88],"individual":[89],"samples.":[90],"paper":[92],"studies":[93],"roots":[95],"these":[97],"variances.":[98],"We":[99,121,182,232],"demonstrate":[100],"CEAP,":[102],"our":[103],"improves":[109],"upon":[110],"EAP-IG":[111],"theoretical":[114],"guarantee,":[115],"can":[116,167],"substantially":[117],"lessen":[118],"variance.":[120],"further":[122],"show":[123,183,233],"variance":[126],"arises":[127],"because":[128],"different":[131,136],"templates":[132],"tend":[133],"activate":[135],"circuits":[137],"model.":[140],"leads":[142],"us":[143],"argue":[145,207],"may":[148,176],"be":[149,168,177],"challenging":[150],"find":[152],"comprehensive":[154],"explains":[157],"controls":[159],"model's":[161],"behavior":[162],"task,":[165],"which":[166,186],"expressed":[169],"countless":[171],"templates,":[172],"suggesting":[173],"LLMs":[175],"inherently":[178],"hard":[179],"steer.":[181],"sparsity,":[185],"has":[187],"been":[188],"claimed":[189],"form":[191],"more":[192],"compact":[193],"interpretable":[195],"task":[196],"circuits,":[197],"fails":[198],"solve":[200],"this":[201],"problem.":[202],"Regarding":[203],"largely":[211],"benign:":[212],"extremely":[213,252],"poor":[214,253],"scores":[216,254],"often":[217],"stem":[218],"how":[220],"defined,":[223],"rather":[224],"than":[225],"defects":[227],"measured":[230],"circuits.":[231],"magnitude":[236],"affected":[240],"by":[241],"selective":[242],"contribution":[243],"scaling,":[244],"neural":[246],"mechanism":[247],"accounts":[249],"sometimes":[255],"observed.":[256]},"counts_by_year":[],"updated_date":"2026-07-01T06:00:48.157686","created_date":"2026-06-17T00:00:00"}