{"authors":[{"id":"orcid_______::1d759d4f48892e627622432a84b2b801","fullName":"Piras, Giorgio","name":"Giorgio","surname":"Piras","rank":1,"pid":{"id":{"scheme":"orcid","value":"0000-0001-8225-6138"},"provenance":null}},{"id":"orcid_______::6e133cd29a2e9d1197ae7c336416823f","fullName":"Mura, Raffaele","name":"Raffaele","surname":"Mura","rank":2,"pid":{"id":{"scheme":"orcid","value":"0009-0005-8719-3048"},"provenance":null}},{"id":null,"fullName":"Brau, Fabio","name":"Fabio","surname":"Brau","rank":3,"pid":{"id":{"scheme":"orcid_pending","value":"0000-0002-9948-111x"},"provenance":null}},{"id":"orcid_______::47ac2583ee1f2f0afdd7e82616b75fb0","fullName":"Oneto, Luca","name":"Luca","surname":"Oneto","rank":4,"pid":{"id":{"scheme":"orcid","value":"0000-0002-8445-395x"},"provenance":null}},{"id":null,"fullName":"Roli, Fabio","name":"Fabio","surname":"Roli","rank":5,"pid":{"id":{"scheme":"orcid_pending","value":"0000-0003-4103-9190"},"provenance":null}},{"id":null,"fullName":"Biggio, Battista","name":"Battista","surname":"Biggio","rank":6,"pid":{"id":{"scheme":"orcid_pending","value":"0000-0001-7752-509x"},"provenance":null}}],"openAccessColor":null,"publiclyFunded":false,"eoscIfGuidelines":null,"type":"publication","language":{"code":"und","label":"Undetermined"},"countries":[{"code":"IT","label":"Italy","provenance":null}],"subjects":[{"subject":{"scheme":"keyword","value":"Machine Learning"},"provenance":null},{"subject":{"scheme":"keyword","value":"FOS: Computer and information sciences"},"provenance":null},{"subject":{"scheme":"keyword","value":"Artificial Intelligence (cs.AI)"},"provenance":null},{"subject":{"scheme":"keyword","value":"Artificial Intelligence"},"provenance":null},{"subject":{"scheme":"FOS","value":"0202 electrical engineering, electronic engineering, information engineering"},"provenance":null},{"subject":{"scheme":"FOS","value":"02 engineering and technology"},"provenance":null},{"subject":{"scheme":"keyword","value":"Machine Learning (cs.LG)"},"provenance":null}],"mainTitle":"SOM Directions Are Better than One: Multi-Directional Refusal Suppression in Language Models","subTitle":null,"descriptions":["<jats:p>Refusal refers to the functional behavior enabling safety-aligned language models to reject harmful or unethical prompts. Following the growing scientific interest in mechanistic interpretability, recent work encoded refusal behavior as a single direction in the model’s latent space; e.g., computed as the difference between the centroids of harmful and harmless prompt representations. However, emerging evidence suggests that concepts in LLMs often appear to be encoded as a low-dimensional manifold embedded in the high-dimensional latent space. Motivated by these findings, we propose a novel method leveraging Self-Organizing Maps (SOMs) to extract multiple refusal directions. To this end, we first prove that SOMs generalize the prior work's difference-in-means technique. We then train SOMs on harmful prompt representations to identify multiple neurons. By subtracting the centroid of harmless representations from each neuron, we derive a set of multiple directions expressing the refusal concept. We validate our method on an extensive experimental setup, demonstrating that ablating multiple directions from models' internals outperforms not only the single-direction baseline but also specialized jailbreak algorithms, leading to an effective suppression of refusal. Finally, we conclude by analyzing the mechanistic implications of our approach.</jats:p>"],"publicationDate":"2026-03-14","publisher":"Association for the Advancement of Artificial Intelligence (AAAI)","embargoEndDate":"2025-11-01","sources":["Crossref"],"formats":["application/pdf"],"contributors":null,"coverages":null,"bestAccessRight":{"code":"c_abf2","label":"OPEN","scheme":"http://vocabularies.coar-repositories.org/documentation/access_rights/"},"container":{"name":"Proceedings of the AAAI Conference on Artificial Intelligence","issnPrinted":"2159-5399","issnOnline":"2374-3468","issnLinking":null,"ep":"32736","iss":null,"sp":"32728","vol":"40","edition":null,"conferencePlace":null,"conferenceDate":null},"documentationUrls":null,"codeRepositoryUrl":null,"programmingLanguage":null,"contactPeople":null,"contactGroups":null,"tools":null,"size":null,"version":null,"geoLocations":null,"id":"doi_dedup___::a28adcd9475ca2518e6d41e06ac4e630","originalIds":["10.1609/aaai.v40i39.40551","50|doiboost____|b1abb2c2f7358aeb24a0b28b09ab4583","50|datacite____::a28adcd9475ca2518e6d41e06ac4e630","10.48550/arxiv.2511.08379","50|od________18::f98f4a66e26bf9713336bb73de68d31e","oai:arXiv.org:2511.08379","50|dblp________::02ebfd519a674fca52299674cbcb2451","50|dblp________::7640280ea7cc124aa04745c38d3d95b0","50|od______3705::5bdc3e1ff365673c200a276e33e7dc56","oai:iris.unica.it:11584/476645","50|od______3710::c547eed678c43ebbe5822113526606a7","oai:unige.iris.cineca.it:11567/1297266"],"pids":[{"scheme":"doi","value":"10.1609/aaai.v40i39.40551"},{"scheme":"doi","value":"10.48550/arxiv.2511.08379"},{"scheme":"arXiv","value":"2511.08379"},{"scheme":"handle","value":"11584/476645"},{"scheme":"handle","value":"11567/1297266"}],"dateOfCollection":null,"lastUpdateTimeStamp":null,"indicators":{"citationImpact":{"citationCount":0.0,"influence":2.2251732E-9,"popularity":2.7165745E-9,"impulse":0.0,"citationClass":"C5","influenceClass":"C5","impulseClass":"C5","popularityClass":"C5"}},"projects":[{"id":"corda_____he::1c53907b78bb7a7765ce158190572b73","code":"101120393","acronym":"Sec4AI4Sec","title":"Cybersecurity for AI-Augmented Systems","funder":"European Commission","pids":[{"scheme":"doi","value":"10.3030/101120393"}]},{"id":"corda_____he::11de478d3259cb6c919efec4eea8ae9b","code":"101070617","acronym":"ELSA","title":"European Lighthouse on Secure and Safe AI","funder":"European Commission","pids":[{"scheme":"doi","value":"10.3030/101070617"}]},{"id":"corda_____he::c2ec719abcb36ba05aa7bddb901c59a8","code":"101168560","acronym":"CoEvolution","title":"A COMPREHENSIVE TRUSTWORTHY FRAMEWORK FOR CONNECTED MACHINE LEARNING AND SECURE INTERCONNECTED AI SOLUTIONS","funder":"European Commission","pids":[{"scheme":"doi","value":"10.3030/101168560"}]}],"organizations":[{"legalName":"University of Genoa","acronym":"UniGe","id":"openorgs____::53df715e03d652944df3d336afe1181a","pids":[{"scheme":"PIC","value":"999976687"},{"scheme":"OrgReg","value":"IT0028"},{"scheme":"ISNI","value":"0000000121513065"},{"scheme":"Wikidata","value":"Q593321"},{"scheme":"ROR","value":"https://ror.org/0107c5v14"},{"scheme":"FundRef","value":"501100004702"},{"scheme":"mag_id","value":"83816512"},{"scheme":"fundref","value":"501100004702"},{"scheme":"wikidata","value":"Q593321"},{"scheme":"GRID","value":"grid.5606.5"},{"scheme":"OrgRef","value":"456119"}]},{"legalName":"University of Cagliari","acronym":"University of Cagliari","id":"openorgs____::9be673852bd7263f5619812856adc52a","pids":[{"scheme":"Wikidata","value":"Q39918611"},{"scheme":"fundref","value":"501100002772"},{"scheme":"OrgReg","value":"IT0013"},{"scheme":"wikidata","value":"Q39918611"},{"scheme":"PIC","value":"999841663"},{"scheme":"GRID","value":"grid.7763.5"},{"scheme":"wikidata","value":"Q2067315"},{"scheme":"ISNI","value":"0000000417553242"},{"scheme":"OrgRef","value":"4958826"},{"scheme":"fundref","value":"100013003"},{"scheme":"Wikidata","value":"Q2067315"},{"scheme":"ROR","value":"https://ror.org/003109y17"},{"scheme":"FundRef","value":"100013003"},{"scheme":"mag_id","value":"172446870"},{"scheme":"FundRef","value":"501100002772"}]}],"communities":null,"collectedFrom":[{"key":"openaire____::9e3be59865b2c1c335d32dae2fe7b254","value":"Datacite"},{"key":"opendoar____::2cfa3753d6a524711acb5fce38eeca1a","value":"Archivio istituzionale della ricerca - Università di Cagliari"},{"key":"openaire____::081b82f96300b6a6e3d282bad31cb6e2","value":"Crossref"},{"key":"openaire____::d8b68cc0a53121f6f896883a7d60c1db","value":"DBLP"},{"key":"opendoar____::5505712229fb1eb500efadddc0353264","value":"Archivio istituzionale della ricerca - Università di Genova"},{"key":"opendoar____::6f4922f45568161a8cdf4ad2299f6d23","value":"arXiv.org e-Print Archive"}],"instances":[{"pids":[{"scheme":"doi","value":"10.1609/aaai.v40i39.40551"}],"type":"Article","urls":["https://doi.org/10.1609/aaai.v40i39.40551"],"publicationDate":"2026-03-14","refereed":"peerReviewed","hostedBy":{"key":"issn___print::c3eeaf213793b1a8aa96af199fd54aa4","value":"Proceedings of the AAAI Conference on Artificial Intelligence"},"collectedFrom":{"key":"openaire____::081b82f96300b6a6e3d282bad31cb6e2","value":"Crossref"}},{"pids":[{"scheme":"doi","value":"10.48550/arxiv.2511.08379"}],"license":"CC BY","type":"Article","urls":["https://dx.doi.org/10.48550/arxiv.2511.08379"],"publicationDate":"2025-01-01","refereed":"nonPeerReviewed","hostedBy":{"key":"openaire____::55045bd2a65019fd8e6741a755395c8c","value":"Unknown Repository"},"collectedFrom":{"key":"openaire____::9e3be59865b2c1c335d32dae2fe7b254","value":"Datacite"}},{"pids":[{"scheme":"arXiv","value":"2511.08379"}],"alternateIdentifiers":[{"scheme":"doi","value":"10.1609/aaai.v40i39.40551"}],"accessRight":{"code":"c_abf2","label":"OPEN","scheme":"http://vocabularies.coar-repositories.org/documentation/access_rights/","openAccessRoute":null},"type":"Preprint","urls":["https://arxiv.org/abs/2511.08379","http://arxiv.org/abs/2511.08379"],"publicationDate":"2025-11-11","refereed":"nonPeerReviewed","hostedBy":{"key":"opendoar____::6f4922f45568161a8cdf4ad2299f6d23","value":"arXiv.org e-Print Archive"},"collectedFrom":{"key":"opendoar____::6f4922f45568161a8cdf4ad2299f6d23","value":"arXiv.org e-Print Archive"}},{"alternateIdentifiers":[{"scheme":"doi","value":"10.48550/arxiv.2511.08379"}],"type":"Preprint","urls":["https://doi.org/10.48550/arXiv.2511.08379"],"publicationDate":"2025-01-01","refereed":"nonPeerReviewed","hostedBy":{"key":"openaire____::d8b68cc0a53121f6f896883a7d60c1db","value":"DBLP"},"collectedFrom":{"key":"openaire____::d8b68cc0a53121f6f896883a7d60c1db","value":"DBLP"}},{"alternateIdentifiers":[{"scheme":"doi","value":"10.1609/aaai.v40i39.40551"}],"type":"Conference object","urls":["https://doi.org/10.1609/aaai.v40i39.40551"],"refereed":"nonPeerReviewed","hostedBy":{"key":"openaire____::d8b68cc0a53121f6f896883a7d60c1db","value":"DBLP"},"collectedFrom":{"key":"openaire____::d8b68cc0a53121f6f896883a7d60c1db","value":"DBLP"}},{"pids":[{"scheme":"handle","value":"11584/476645"}],"alternateIdentifiers":[{"scheme":"urn","value":"https://hdl.handle.net/11584/476645"},{"scheme":"doi","value":"10.1609/aaai.v40i39.40551"}],"accessRight":{"code":"c_14cb","label":"CLOSED","scheme":"http://vocabularies.coar-repositories.org/documentation/access_rights/","openAccessRoute":null},"type":"Conference object","urls":["https://hdl.handle.net/11584/476645","https://doi.org/10.1609/aaai.v40i39.40551"],"publicationDate":"2026-01-01","refereed":"nonPeerReviewed","hostedBy":{"key":"opendoar____::2cfa3753d6a524711acb5fce38eeca1a","value":"Archivio istituzionale della ricerca - Università di Cagliari"},"collectedFrom":{"key":"opendoar____::2cfa3753d6a524711acb5fce38eeca1a","value":"Archivio istituzionale della ricerca - Università di Cagliari"}},{"pids":[{"scheme":"handle","value":"11567/1297266"}],"alternateIdentifiers":[{"scheme":"urn","value":"2-s2.0-105034965285"},{"scheme":"doi","value":"10.1609/aaai.v40i39.40551"}],"accessRight":{"code":"c_16ec","label":"RESTRICTED","scheme":"http://vocabularies.coar-repositories.org/documentation/access_rights/","openAccessRoute":null},"type":"Conference object","urls":["https://hdl.handle.net/11567/1297266","https://doi.org/10.1609/aaai.v40i39.40551"],"publicationDate":"2026-01-01","refereed":"nonPeerReviewed","hostedBy":{"key":"opendoar____::5505712229fb1eb500efadddc0353264","value":"Archivio istituzionale della ricerca - Università di Genova"},"collectedFrom":{"key":"opendoar____::5505712229fb1eb500efadddc0353264","value":"Archivio istituzionale della ricerca - Università di Genova"}}],"links":[{"header":{"relationType":"resultProject","relationClass":"isProducedBy","relatedIdentifier":"corda_____he::1c53907b78bb7a7765ce158190572b73","relatedRecordType":"project","relationProvenance":"sysimport:crosswalk:repository","trust":"0.9"},"collectedfrom":[{"dsId":"openaire____::3f264f93cf3b0cfc4ede188a6300455c","dsName":"CORDA - COmmon Research DAta Warehouse - Horizon Europe"}],"projectTitle":"Cybersecurity for AI-Augmented Systems","code":"101120393","funding":{"funder":{"id":"ec__________::EC","shortname":"EC","name":"European Commission","jurisdiction":{"code":"EU","label":"European Union"},"pid":null},"level0":{"id":"ec__________::EC::HE","description":"Horizon Europe Framework Programme","name":"HE"},"level1":{"id":"ec__________::EC::HE::HORIZON-RIA","description":"HORIZON  Research and Innovation Actions","name":"HORIZON-RIA"},"level2":{"id":null,"description":null,"name":null}},"startDate":"2023-10-01","endDate":"2026-09-30"},{"header":{"relationType":"resultProject","relationClass":"isProducedBy","relatedIdentifier":"corda_____he::11de478d3259cb6c919efec4eea8ae9b","relatedRecordType":"project","relationProvenance":"sysimport:crosswalk:repository","trust":"0.9"},"collectedfrom":[{"dsId":"openaire____::3f264f93cf3b0cfc4ede188a6300455c","dsName":"CORDA - COmmon Research DAta Warehouse - Horizon Europe"}],"projectTitle":"European Lighthouse on Secure and Safe AI","code":"101070617","funding":{"funder":{"id":"ec__________::EC","shortname":"EC","name":"European Commission","jurisdiction":{"code":"EU","label":"European Union"},"pid":null},"level0":{"id":"ec__________::EC::HE","description":"Horizon Europe Framework Programme","name":"HE"},"level1":{"id":"ec__________::EC::HE::HORIZON-RIA","description":"HORIZON  Research and Innovation Actions","name":"HORIZON-RIA"},"level2":{"id":null,"description":null,"name":null}},"startDate":"2022-09-01","endDate":"2026-08-31"},{"header":{"relationType":"resultProject","relationClass":"isProducedBy","relatedIdentifier":"corda_____he::c2ec719abcb36ba05aa7bddb901c59a8","relatedRecordType":"project","relationProvenance":"sysimport:crosswalk:repository","trust":"0.9"},"collectedfrom":[{"dsId":"openaire____::3f264f93cf3b0cfc4ede188a6300455c","dsName":"CORDA - COmmon Research DAta Warehouse - Horizon Europe"}],"projectTitle":"A COMPREHENSIVE TRUSTWORTHY FRAMEWORK FOR CONNECTED MACHINE LEARNING AND SECURE INTERCONNECTED AI SOLUTIONS","code":"101168560","funding":{"funder":{"id":"ec__________::EC","shortname":"EC","name":"European Commission","jurisdiction":{"code":"EU","label":"European Union"},"pid":null},"level0":{"id":"ec__________::EC::HE","description":"Horizon Europe Framework Programme","name":"HE"},"level1":{"id":"ec__________::EC::HE::HORIZON-RIA","description":"HORIZON  Research and Innovation Actions","name":"HORIZON-RIA"},"level2":{"id":null,"description":null,"name":null}},"startDate":"2024-11-01","endDate":"2027-10-31"}],"otherTitles":null,"inDiamondJournal":false,"green":true,"isGreen":true,"isInDiamondJournal":false}