{"id":"https://openalex.org/W7137881767","doi":"https://doi.org/10.48550/arxiv.2603.13318","title":"Residual Stream Analysis of Overfitting And Structural Disruptions","display_name":"Residual Stream Analysis of Overfitting And Structural Disruptions","publication_year":2026,"publication_date":"2026-03-04","ids":{"openalex":"https://openalex.org/W7137881767","doi":"https://doi.org/10.48550/arxiv.2603.13318"},"language":null,"primary_location":{"id":"doi:10.48550/arxiv.2603.13318","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2603.13318","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":"cc-by","license_id":"https://openalex.org/licenses/cc-by","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.2603.13318","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5129651347","display_name":"Quan Liu","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Liu, Quan","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5129691480","display_name":"Han Zhou","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Zhou, Han","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5129739172","display_name":"Wenquan Wu","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Wu, Wenquan","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5129678029","display_name":"Hua Wu","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Wu, Hua","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5129645671","display_name":"Sen Su","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Su, Sen","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/T10028","display_name":"Topic Modeling","score":0.26750001311302185,"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/T10028","display_name":"Topic Modeling","score":0.26750001311302185,"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":0.22529999911785126,"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/T11689","display_name":"Adversarial Robustness in Machine Learning","score":0.11959999799728394,"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/overfitting","display_name":"Overfitting","score":0.8705999851226807},{"id":"https://openalex.org/keywords/residual","display_name":"Residual","score":0.7132999897003174},{"id":"https://openalex.org/keywords/variance","display_name":"Variance (accounting)","score":0.6371999979019165},{"id":"https://openalex.org/keywords/entropy","display_name":"Entropy (arrow of time)","score":0.5601000189781189},{"id":"https://openalex.org/keywords/smoothness","display_name":"Smoothness","score":0.5293999910354614},{"id":"https://openalex.org/keywords/security-token","display_name":"Security token","score":0.38119998574256897},{"id":"https://openalex.org/keywords/overhead","display_name":"Overhead (engineering)","score":0.3278000056743622}],"concepts":[{"id":"https://openalex.org/C22019652","wikidata":"https://www.wikidata.org/wiki/Q331309","display_name":"Overfitting","level":3,"score":0.8705999851226807},{"id":"https://openalex.org/C155512373","wikidata":"https://www.wikidata.org/wiki/Q287450","display_name":"Residual","level":2,"score":0.7132999897003174},{"id":"https://openalex.org/C196083921","wikidata":"https://www.wikidata.org/wiki/Q7915758","display_name":"Variance (accounting)","level":2,"score":0.6371999979019165},{"id":"https://openalex.org/C106301342","wikidata":"https://www.wikidata.org/wiki/Q4117933","display_name":"Entropy (arrow of time)","level":2,"score":0.5601000189781189},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.5343000292778015},{"id":"https://openalex.org/C102634674","wikidata":"https://www.wikidata.org/wiki/Q868473","display_name":"Smoothness","level":2,"score":0.5293999910354614},{"id":"https://openalex.org/C105795698","wikidata":"https://www.wikidata.org/wiki/Q12483","display_name":"Statistics","level":1,"score":0.5250999927520752},{"id":"https://openalex.org/C33923547","wikidata":"https://www.wikidata.org/wiki/Q395","display_name":"Mathematics","level":0,"score":0.4478999972343445},{"id":"https://openalex.org/C48145219","wikidata":"https://www.wikidata.org/wiki/Q1335365","display_name":"Security token","level":2,"score":0.38119998574256897},{"id":"https://openalex.org/C124101348","wikidata":"https://www.wikidata.org/wiki/Q172491","display_name":"Data mining","level":1,"score":0.3294999897480011},{"id":"https://openalex.org/C2779960059","wikidata":"https://www.wikidata.org/wiki/Q7113681","display_name":"Overhead (engineering)","level":2,"score":0.3278000056743622},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.30239999294281006},{"id":"https://openalex.org/C175291020","wikidata":"https://www.wikidata.org/wiki/Q1156822","display_name":"Offset (computer science)","level":2,"score":0.29089999198913574},{"id":"https://openalex.org/C11413529","wikidata":"https://www.wikidata.org/wiki/Q8366","display_name":"Algorithm","level":1,"score":0.2858000099658966},{"id":"https://openalex.org/C3018076075","wikidata":"https://www.wikidata.org/wiki/Q1826427","display_name":"Variance components","level":2,"score":0.2858000099658966},{"id":"https://openalex.org/C149782125","wikidata":"https://www.wikidata.org/wiki/Q160039","display_name":"Econometrics","level":1,"score":0.28279998898506165},{"id":"https://openalex.org/C81917197","wikidata":"https://www.wikidata.org/wiki/Q628760","display_name":"Selection (genetic algorithm)","level":2,"score":0.2773999869823456},{"id":"https://openalex.org/C153180895","wikidata":"https://www.wikidata.org/wiki/Q7148389","display_name":"Pattern recognition (psychology)","level":2,"score":0.27239999175071716},{"id":"https://openalex.org/C56951928","wikidata":"https://www.wikidata.org/wiki/Q3539213","display_name":"Trimming","level":2,"score":0.26109999418258667},{"id":"https://openalex.org/C140779682","wikidata":"https://www.wikidata.org/wiki/Q210868","display_name":"Sampling (signal processing)","level":3,"score":0.25949999690055847}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.48550/arxiv.2603.13318","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2603.13318","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":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":null,"is_accepted":false,"is_published":null,"raw_source_name":null,"raw_type":"Preprint"}],"best_oa_location":{"id":"doi:10.48550/arxiv.2603.13318","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2603.13318","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":"cc-by","license_id":"https://openalex.org/licenses/cc-by","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":{"Ensuring":[0],"that":[1,46,81,133,144],"large":[2],"language":[3],"models":[4],"(LLMs)":[5],"remain":[6],"both":[7],"helpful":[8],"and":[9,54,79,96,165],"harmless":[10],"poses":[11],"a":[12,71,90],"significant":[13],"challenge:":[14],"fine-tuning":[15],"on":[16,159],"repetitive":[17],"safety":[18,47,85,111],"datasets,":[19],"where":[20],"unsafe":[21],"prompts":[22],"are":[23,38],"paired":[24],"with":[25],"standard":[26],"refusal":[27,101],"templates,":[28],"often":[29],"leads":[30],"to":[31,59,107,117],"false":[32,98,147],"refusals,":[33],"in":[34,138],"which":[35],"benign":[36],"queries":[37],"declined.":[39],"We":[40],"first":[41],"quantify":[42],"this":[43],"effect,":[44],"showing":[45],"data":[48,112],"exhibits":[49],"substantially":[50],"lower":[51],"token":[52],"entropy":[53],"2-gram":[55],"diversity":[56],"(0.048)":[57],"compared":[58],"general":[60,160],"instruction":[61],"data.":[62],"To":[63],"uncover":[64],"the":[65],"root":[66],"cause,":[67],"we":[68,124],"introduce":[69],"FlowLens,":[70],"stable":[72],"PCA-based":[73],"tool":[74],"for":[75],"residual-stream":[76],"geometry":[77],"analysis,":[78],"reveal":[80],"higher":[82],"proportions":[83],"of":[84],"examples":[86],"concentrate":[87],"variance":[88,136],"along":[89],"few":[91],"components,":[92],"reducing":[93],"representational":[94],"smoothness":[95],"driving":[97],"refusals":[99,148],"(false":[100],"rate":[102],"rises":[103],"from":[104,114],"63":[105],"percent":[106,109,116],"84":[108],"as":[110,163],"increases":[113],"0":[115],"40":[118],"percent).":[119],"Guided":[120],"by":[121,149],"these":[122],"insights,":[123],"propose":[125],"Variance":[126],"Concentration":[127],"Loss":[128],"(VCL),":[129],"an":[130],"auxiliary":[131],"regularizer":[132],"penalizes":[134],"excessive":[135],"concentration":[137],"mid-layer":[139],"residuals.":[140],"Empirical":[141],"results":[142],"demonstrate":[143],"VCL":[145],"reduces":[146],"over":[150],"35":[151],"percentage":[152],"points":[153],"while":[154],"maintaining":[155],"or":[156],"improving":[157],"performance":[158],"benchmarks":[161],"such":[162],"MMLU":[164],"GSM8K.":[166]},"counts_by_year":[],"updated_date":"2026-07-01T06:00:48.157686","created_date":"2026-03-18T00:00:00"}