{"id":"https://openalex.org/W7172294263","doi":"https://doi.org/10.48550/arxiv.2607.28969","title":"SafeNexus: Discovering and Steering Modality-Universal Safety Neurons in MLLMs","display_name":"SafeNexus: Discovering and Steering Modality-Universal Safety Neurons in MLLMs","publication_year":2026,"publication_date":"2026-07-31","ids":{"openalex":"https://openalex.org/W7172294263","doi":"https://doi.org/10.48550/arxiv.2607.28969"},"language":null,"primary_location":{"id":"doi:10.48550/arxiv.2607.28969","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2607.28969","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":"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.2607.28969","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5144258594","display_name":"Jian Yu","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Yu, Jian","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5144305150","display_name":"Fei Shen","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Shen, Fei","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5144281061","display_name":"Cong Wang","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Wang, Cong","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5144258734","display_name":"Jian Wang","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Wang, Jian","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5147054068","display_name":"Lu Jin","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Jin, Lu","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5146941240","display_name":"Xiaoyu Du","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Du, Xiaoyu","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5147053273","display_name":"Jinhui Tang","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Tang, Jinhui","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5146978871","display_name":"Tat-Seng Chua","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Chua, Tat-Seng","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/T11689","display_name":"Adversarial Robustness in Machine Learning","score":0.9739000201225281,"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":0.9739000201225281,"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":0.005499999970197678,"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/T12026","display_name":"Explainable Artificial Intelligence (XAI)","score":0.0026000000070780516,"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/robustness","display_name":"Robustness (evolution)","score":0.5342000126838684},{"id":"https://openalex.org/keywords/limiting","display_name":"Limiting","score":0.5123000144958496},{"id":"https://openalex.org/keywords/salience","display_name":"Salience (neuroscience)","score":0.46860000491142273},{"id":"https://openalex.org/keywords/modality","display_name":"Modality (human\u2013computer interaction)","score":0.42160001397132874},{"id":"https://openalex.org/keywords/exploit","display_name":"Exploit","score":0.4047999978065491},{"id":"https://openalex.org/keywords/modalities","display_name":"Modalities","score":0.3181999921798706},{"id":"https://openalex.org/keywords/bridging","display_name":"Bridging (networking)","score":0.30070000886917114}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.6899999976158142},{"id":"https://openalex.org/C63479239","wikidata":"https://www.wikidata.org/wiki/Q7353546","display_name":"Robustness (evolution)","level":3,"score":0.5342000126838684},{"id":"https://openalex.org/C188198153","wikidata":"https://www.wikidata.org/wiki/Q1613840","display_name":"Limiting","level":2,"score":0.5123000144958496},{"id":"https://openalex.org/C108154423","wikidata":"https://www.wikidata.org/wiki/Q1469792","display_name":"Salience (neuroscience)","level":2,"score":0.46860000491142273},{"id":"https://openalex.org/C2780226545","wikidata":"https://www.wikidata.org/wiki/Q6888030","display_name":"Modality (human\u2013computer interaction)","level":2,"score":0.42160001397132874},{"id":"https://openalex.org/C165696696","wikidata":"https://www.wikidata.org/wiki/Q11287","display_name":"Exploit","level":2,"score":0.4047999978065491},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.3343000113964081},{"id":"https://openalex.org/C2779903281","wikidata":"https://www.wikidata.org/wiki/Q6888026","display_name":"Modalities","level":2,"score":0.3181999921798706},{"id":"https://openalex.org/C174348530","wikidata":"https://www.wikidata.org/wiki/Q188635","display_name":"Bridging (networking)","level":2,"score":0.30070000886917114},{"id":"https://openalex.org/C112930515","wikidata":"https://www.wikidata.org/wiki/Q4389547","display_name":"Risk analysis (engineering)","level":1,"score":0.27810001373291016},{"id":"https://openalex.org/C3017944768","wikidata":"https://www.wikidata.org/wiki/Q1450463","display_name":"Poison control","level":2,"score":0.27309998869895935},{"id":"https://openalex.org/C100776233","wikidata":"https://www.wikidata.org/wiki/Q2532492","display_name":"Bridge (graph theory)","level":2,"score":0.2720000147819519},{"id":"https://openalex.org/C148493468","wikidata":"https://www.wikidata.org/wiki/Q2646951","display_name":"Functional safety","level":2,"score":0.26989999413490295},{"id":"https://openalex.org/C132835097","wikidata":"https://www.wikidata.org/wiki/Q7663745","display_name":"System safety","level":2,"score":0.26170000433921814},{"id":"https://openalex.org/C26760741","wikidata":"https://www.wikidata.org/wiki/Q160402","display_name":"Perception","level":2,"score":0.2524000108242035},{"id":"https://openalex.org/C48044578","wikidata":"https://www.wikidata.org/wiki/Q727490","display_name":"Scalability","level":2,"score":0.2506999969482422}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.48550/arxiv.2607.28969","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2607.28969","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":"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.2607.28969","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2607.28969","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":"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":{"Although":[0],"Large":[1,14],"Language":[2,15],"Models":[3,16],"(LLMs)":[4],"have":[5],"demonstrated":[6],"promising":[7],"safety":[8,28,56,102,111,124,155,172,176,179,187,220],"performance,":[9],"extending":[10],"them":[11,205],"to":[12,35,99],"Multimodal":[13],"(MLLMs)":[17],"exposes":[18],"a":[19,54,61,69],"significant":[20],"gap":[21],"between":[22],"expanded":[23],"multimodal":[24],"capabilities":[25],"and":[26,83,105,178],"existing":[27],"mechanisms.":[29],"Current":[30],"defenses":[31],"remain":[32],"predominantly":[33],"confined":[34],"specific":[36],"modal":[37],"settings,":[38],"thereby":[39],"limiting":[40],"their":[41,85,107],"robustness":[42],"against":[43,143],"broader":[44],"cross-modal":[45,55,120,145],"threats.":[46],"To":[47],"bridge":[48],"this":[49,93],"gap,":[50],"we":[51,67,95,169],"introduce":[52],"SafeNexus,":[53],"alignment":[57,173],"framework":[58],"that":[59,73,149,212],"adopts":[60],"dedicated":[62],"neuron-level":[63],"intervention":[64],"strategy.":[65],"First,":[66],"formulate":[68],"neuron":[70,180],"localization":[71],"paradigm":[72],"identifies":[74],"functionally":[75],"specialized":[76],"neurons":[77,103,125,152],"by":[78],"characterizing":[79],"intermediate-layer":[80],"activation":[81,196],"patterns":[82],"quantifying":[84],"functional":[86],"salience":[87],"through":[88,188],"importance":[89],"scoring.":[90],"Building":[91,165],"upon":[92],"paradigm,":[94],"exploit":[96],"contrastive":[97],"data":[98],"identify":[100],"modality-bound":[101],"(BS-Neurons),":[104],"validate":[106],"role":[108],"in":[109],"regulating":[110],"behavior":[112],"within":[113],"each":[114],"modality":[115,224],"via":[116,206],"targeted":[117,207],"suppression.":[118],"Further":[119],"analysis":[121],"defines":[122],"modality-universal":[123],"(US-Neurons)":[126],"as":[127,138],"the":[128,139,192,195,201],"shared":[129],"subset":[130],"of":[131,198],"BS-Neurons":[132],"identified":[133],"across":[134,157],"individual":[135],"modalities,":[136,158],"serving":[137],"core":[140],"for":[141],"defending":[142],"harmful":[144],"attacks.":[146],"We":[147],"observe":[148],"suppressing":[150],"these":[151,167],"substantially":[153],"degrades":[154],"performance":[156],"while":[159,200,226],"leaving":[160],"overall":[161],"utility":[162],"largely":[163],"unaffected.":[164],"on":[166,219],"insights,":[168],"propose":[170],"two":[171,189],"strategies:":[174],"activation-level":[175],"amplifier":[177],"calibrator.":[181],"The":[182],"proposed":[183],"strategies":[184],"enhance":[185],"model":[186],"distinct":[190],"routes:":[191],"former":[193],"amplifies":[194],"magnitudes":[197],"US-Neurons,":[199],"latter":[202],"selectively":[203],"calibrates":[204],"fine-tuning.":[208],"Extensive":[209],"experiments":[210],"demonstrate":[211],"our":[213],"method":[214],"outperforms":[215],"prevailing":[216],"state-of-the-art":[217],"approaches":[218],"benchmarks":[221],"spanning":[222],"diverse":[223],"combinations,":[225],"effectively":[227],"preserving":[228],"utility.":[229]},"counts_by_year":[],"updated_date":"2026-08-13T07:04:57.449891","created_date":"2026-08-04T00:00:00"}