{"id":"https://openalex.org/W7130676367","doi":"https://doi.org/10.48550/arxiv.2602.16835","title":"NeST: Neuron Selective Tuning for LLM Safety","display_name":"NeST: Neuron Selective Tuning for LLM Safety","publication_year":2026,"publication_date":"2026-02-18","ids":{"openalex":"https://openalex.org/W7130676367","doi":"https://doi.org/10.48550/arxiv.2602.16835"},"language":"en","primary_location":{"id":"pmh:oai:arXiv.org:2602.16835","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2602.16835","pdf_url":"https://arxiv.org/pdf/2602.16835","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":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},"volume":null,"type":"preprint","is_traditional":true,"indexed_in":["arxiv","datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://arxiv.org/pdf/2602.16835","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5121354056","display_name":"Sasha Behrouzi","orcid":"https://orcid.org/0000-0002-7938-9818","observed_orcids":["https://orcid.org/0000-0002-7938-9818"]},"institutions":[{"id":"https://openalex.org/I31512782","display_name":"Technische Universit\u00e4t Darmstadt","ror":"https://ror.org/05n911h24","country_code":"DE","type":"education","lineage":["https://openalex.org/I31512782"]}],"countries":["DE"],"is_corresponding":false,"raw_author_name":"Behrouzi, Sasha","raw_affiliation_strings":["Technical University of Darmstadt"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Technical University of Darmstadt","institution_ids":["https://openalex.org/I31512782"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5045570758","display_name":"Lichao Wu","orcid":"https://orcid.org/0000-0002-7139-732X","observed_orcids":["https://orcid.org/0000-0002-7139-732X"]},"institutions":[{"id":"https://openalex.org/I36234482","display_name":"University of Bristol","ror":"https://ror.org/0524sp257","country_code":"GB","type":"education","lineage":["https://openalex.org/I36234482"]}],"countries":["GB"],"is_corresponding":false,"raw_author_name":"Wu, Lichao","raw_affiliation_strings":["University of Bristol"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"University of Bristol","institution_ids":["https://openalex.org/I36234482"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5090473338","display_name":"Mohamadreza Rostami","orcid":"https://orcid.org/0009-0005-7293-2733","observed_orcids":["https://orcid.org/0009-0005-7293-2733"]},"institutions":[{"id":"https://openalex.org/I31512782","display_name":"Technische Universit\u00e4t Darmstadt","ror":"https://ror.org/05n911h24","country_code":"DE","type":"education","lineage":["https://openalex.org/I31512782"]}],"countries":["DE"],"is_corresponding":false,"raw_author_name":"Rostami, Mohamadreza","raw_affiliation_strings":["Technical University of Darmstadt"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Technical University of Darmstadt","institution_ids":["https://openalex.org/I31512782"]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5121595117","display_name":"Ahmad-Reza Sadeghi","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I31512782","display_name":"Technische Universit\u00e4t Darmstadt","ror":"https://ror.org/05n911h24","country_code":"DE","type":"education","lineage":["https://openalex.org/I31512782"]}],"countries":["DE"],"is_corresponding":false,"raw_author_name":"Sadeghi, Ahmad-Reza","raw_affiliation_strings":["Technical University of Darmstadt"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Technical University of Darmstadt","institution_ids":["https://openalex.org/I31512782"]}]}],"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":3,"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/T11689","display_name":"Adversarial Robustness in Machine Learning","score":0.9984999895095825,"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/T11689","display_name":"Adversarial Robustness in Machine Learning","score":0.9984999895095825,"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":0.00019999999494757503,"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/T12262","display_name":"Hate Speech and Cyberbullying Detection","score":9.999999747378752e-05,"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-model-safety","display_name":"large language model safety","score":0.8610000014305115},{"id":"https://openalex.org/keywords/safety-alignment","display_name":"safety alignment","score":0.7889999747276306},{"id":"https://openalex.org/keywords/jailbreak-attacks","display_name":"jailbreak attacks","score":0.7279999852180481},{"id":"https://openalex.org/keywords/parameter-efficient-fine-tuning","display_name":"parameter-efficient fine-tuning","score":0.7979999780654907},{"id":"https://openalex.org/keywords/multimodal-models","display_name":"multimodal models","score":0.6029999852180481},{"id":"https://openalex.org/keywords/open-weight-models","display_name":"open-weight models","score":0.5770000219345093}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.6545000076293945},{"id":"https://openalex.org/C185798385","wikidata":"https://www.wikidata.org/wiki/Q1161707","display_name":"Benchmark (surveying)","level":2,"score":0.6107000112533569},{"id":"https://openalex.org/C105339364","wikidata":"https://www.wikidata.org/wiki/Q2297740","display_name":"Software deployment","level":2,"score":0.536300003528595},{"id":"https://openalex.org/C21200559","wikidata":"https://www.wikidata.org/wiki/Q7451068","display_name":"Sensitivity (control systems)","level":2,"score":0.4242999851703644},{"id":"https://openalex.org/C111335779","wikidata":"https://www.wikidata.org/wiki/Q3454686","display_name":"Reduction (mathematics)","level":2,"score":0.41280001401901245},{"id":"https://openalex.org/C139807058","wikidata":"https://www.wikidata.org/wiki/Q352374","display_name":"Adaptation (eye)","level":2,"score":0.3978999853134155},{"id":"https://openalex.org/C3017944768","wikidata":"https://www.wikidata.org/wiki/Q1450463","display_name":"Poison control","level":2,"score":0.35190001130104065},{"id":"https://openalex.org/C2780864053","wikidata":"https://www.wikidata.org/wiki/Q5147495","display_name":"Collision avoidance","level":3,"score":0.34790000319480896},{"id":"https://openalex.org/C132835097","wikidata":"https://www.wikidata.org/wiki/Q7663745","display_name":"System safety","level":2,"score":0.33169999718666077},{"id":"https://openalex.org/C168167062","wikidata":"https://www.wikidata.org/wiki/Q1117970","display_name":"Component (thermodynamics)","level":2,"score":0.3010999858379364},{"id":"https://openalex.org/C38652104","wikidata":"https://www.wikidata.org/wiki/Q3510521","display_name":"Computer security","level":1,"score":0.28949999809265137},{"id":"https://openalex.org/C199535089","wikidata":"https://www.wikidata.org/wiki/Q1303258","display_name":"Intrinsic safety","level":3,"score":0.28130000829696655},{"id":"https://openalex.org/C2780385302","wikidata":"https://www.wikidata.org/wiki/Q367158","display_name":"Protocol (science)","level":3,"score":0.2806999981403351},{"id":"https://openalex.org/C2780665704","wikidata":"https://www.wikidata.org/wiki/Q959298","display_name":"Intervention (counseling)","level":2,"score":0.2806999981403351},{"id":"https://openalex.org/C26517878","wikidata":"https://www.wikidata.org/wiki/Q228039","display_name":"Key (lock)","level":2,"score":0.27379998564720154},{"id":"https://openalex.org/C112930515","wikidata":"https://www.wikidata.org/wiki/Q4389547","display_name":"Risk analysis (engineering)","level":1,"score":0.27379998564720154},{"id":"https://openalex.org/C200601418","wikidata":"https://www.wikidata.org/wiki/Q2193887","display_name":"Reliability engineering","level":1,"score":0.27239999175071716},{"id":"https://openalex.org/C140547941","wikidata":"https://www.wikidata.org/wiki/Q7797194","display_name":"Threat model","level":2,"score":0.2651999890804291}],"mesh":[],"locations_count":2,"locations":[{"id":"pmh:oai:arXiv.org:2602.16835","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2602.16835","pdf_url":"https://arxiv.org/pdf/2602.16835","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":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},{"id":"doi:10.48550/arxiv.2602.16835","doi":"https://doi.org/10.48550/arxiv.2602.16835","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2602.16835","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":"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"}],"best_oa_location":{"id":"pmh:oai:arXiv.org:2602.16835","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2602.16835","pdf_url":"https://arxiv.org/pdf/2602.16835","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":"cc-by","license_id":"https://openalex.org/licenses/cc-by","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/W7130676367.pdf","grobid_xml":"https://content.openalex.org/works/W7130676367.grobid-xml"},"referenced_works_count":28,"referenced_works":["https://openalex.org/W2794325560","https://openalex.org/W3042667808","https://openalex.org/W3083410900","https://openalex.org/W3206180216","https://openalex.org/W3210904070","https://openalex.org/W4322718191","https://openalex.org/W4327810158","https://openalex.org/W4384920109","https://openalex.org/W4385374425","https://openalex.org/W4387436590","https://openalex.org/W4391766565","https://openalex.org/W4392822465","https://openalex.org/W4394782215","https://openalex.org/W4401306886","https://openalex.org/W4402670895","https://openalex.org/W4402951277","https://openalex.org/W4403746689","https://openalex.org/W4403796339","https://openalex.org/W4405354744","https://openalex.org/W4406231018","https://openalex.org/W4407760043","https://openalex.org/W4414939355","https://openalex.org/W4415082370","https://openalex.org/W4415085804","https://openalex.org/W4417133177","https://openalex.org/W6947991961","https://openalex.org/W7092296780","https://openalex.org/W7160444000"],"related_works":["https://openalex.org/W7128792276","https://openalex.org/W7204271500","https://openalex.org/W6929630730","https://openalex.org/W7140123232","https://openalex.org/W4417521604","https://openalex.org/W4409362484","https://openalex.org/W7203623355","https://openalex.org/W7214520188","https://openalex.org/W7196971478","https://openalex.org/W4414481156"],"abstract_inverted_index":{"Safety":[0,58],"alignment":[1,235],"is":[2,22,128],"essential":[3],"for":[4,49,89,214],"the":[5,75,121,124,153],"responsible":[6],"deployment":[7],"of":[8,123],"Large":[9],"Language":[10],"Models":[11],"(LLMs).":[12],"Yet,":[13],"existing":[14],"approaches":[15,173],"often":[16],"rely":[17],"on":[18,102,131,161,201,242],"heavyweight":[19],"fine-tuning":[20,33,175],"that":[21,78,231],"costly":[23],"to":[24,55,143,194,211,226],"update,":[25],"audit,":[26],"and":[27,37,53,105,114,165,172,213],"maintain":[28],"across":[29],"model":[30,66],"families.":[31],"Full":[32],"incurs":[34],"substantial":[35],"computational":[36],"storage":[38],"overhead,":[39],"while":[40,119,196],"parameter-efficient":[41],"methods,":[42],"e.g.,":[43],"Low-Rank":[44],"Adaptation":[45],"(LoRA),":[46],"trade":[47],"efficiency":[48],"inconsistent":[50],"safety":[51,80,92,220,234,246],"gains":[52],"sensitivity":[54],"design":[56],"choices.":[57],"intervention":[59],"mechanisms":[60],"reduce":[61],"unsafe":[62],"outputs":[63],"without":[64,135],"modifying":[65],"weights,":[67,155],"but":[68],"do":[69],"not":[70],"directly":[71],"shape":[72],"or":[73],"preserve":[74],"internal":[76],"representations":[77],"govern":[79],"behavior.":[81],"We":[82],"present":[83],"NeST,":[84],"a":[85],"Neuron-Selective":[86],"Tuning":[87],"framework":[88],"efficient":[90],"post-hoc":[91],"alignment.":[93],"NeST":[94,127,168,185],"identifies":[95],"safety-relevant":[96],"feed-forward":[97],"neurons":[98,109],"via":[99],"activation":[100,112],"probing":[101],"vanilla":[103,132],"harmful":[104],"benign":[106],"prompts,":[107,134],"clusters":[108],"with":[110,177],"similar":[111],"profiles,":[113],"trains":[115],"shared":[116],"cluster-level":[117],"updates":[118,148],"freezing":[120],"rest":[122],"model.":[125],"Importantly,":[126],"trained":[129],"only":[130,198],"malicious":[133],"using":[136],"jailbreak-specific":[137],"attack":[138,189],"data,":[139],"yet":[140],"generalizes":[141],"robustly":[142],"diverse":[144],"jailbreaks.":[145],"The":[146],"learned":[147],"are":[149],"then":[150],"folded":[151],"into":[152],"original":[154],"incurring":[156],"no":[157],"inference-time":[158],"overhead.":[159],"Evaluated":[160],"14":[162],"open-weight":[163],"language":[164],"multimodal":[166,204],"models,":[167,184],"outperforms":[169],"lightweight":[170],"baselines":[171],"full":[174],"robustness":[176],"significantly":[178],"fewer":[179],"trainable":[180],"parameters.":[181],"On":[182],"text-only":[183],"reduces":[186,207],"average":[187],"jailbreak":[188],"success":[190],"rate":[191],"from":[192,209,224],"44.5%":[193],"1.1%":[195],"training":[197],"0.4M":[199],"parameters":[200],"average.":[202],"Across":[203],"settings,":[205],"it":[206,218],"ASR":[208,223],"55.3%":[210],"1.1%,":[212],"downstream":[215],"fine-tuned":[216],"variants,":[217],"restores":[219],"by":[221,239],"reducing":[222],"53.8%":[225],"0.8%.":[227],"These":[228],"results":[229],"show":[230],"robust,":[232],"maintainable":[233],"can":[236],"be":[237],"achieved":[238],"concentrating":[240],"adaptation":[241],"localized,":[243],"functionally":[244],"coherent":[245],"structures.":[247]},"counts_by_year":[{"year":2026,"cited_by_count":3}],"updated_date":"2026-11-07T11:25:07.637180","created_date":"2026-02-21T00:00:00"}