{"id":"https://openalex.org/W4412944721","doi":"https://doi.org/10.18653/v1/2025.findings-acl.1223","title":"DIESEL: A Lightweight Inference-Time Safety Enhancement for Language Models","display_name":"DIESEL: A Lightweight Inference-Time Safety Enhancement for Language Models","publication_year":2025,"publication_date":"2025-01-01","ids":{"openalex":"https://openalex.org/W4412944721","doi":"https://doi.org/10.18653/v1/2025.findings-acl.1223"},"language":"en","primary_location":{"id":"doi:10.18653/v1/2025.findings-acl.1223","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2025.findings-acl.1223","pdf_url":"https://aclanthology.org/2025.findings-acl.1223.pdf","source":null,"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Findings of the Association for Computational Linguistics: ACL 2025","raw_type":"proceedings-article"},"type":"conference-paper","indexed_in":["crossref"],"open_access":{"is_oa":true,"oa_status":"gold","oa_url":"https://aclanthology.org/2025.findings-acl.1223.pdf","any_repository_has_fulltext":null},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5093744060","display_name":"Ben Ganon","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Ben Ganon","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5033720635","display_name":"Alon Zolfi","orcid":"https://orcid.org/0000-0003-0270-1743"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Alon Zolfi","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5092189525","display_name":"Omer Hofman","orcid":"https://orcid.org/0009-0006-0136-9792"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Omer Hofman","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5101889791","display_name":"Inderjeet Singh","orcid":"https://orcid.org/0000-0003-3011-3199"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Inderjeet Singh","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5077629763","display_name":"Hisashi Kojima","orcid":"https://orcid.org/0000-0002-0973-1673"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Hisashi Kojima","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5072913672","display_name":"Yuval Elovici","orcid":"https://orcid.org/0000-0002-9641-128X"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Yuval Elovici","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5002391103","display_name":"Asaf Shabtai","orcid":"https://orcid.org/0000-0003-0630-4059"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Asaf Shabtai","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":true,"cited_by_count":0,"citation_normalized_percentile":null,"cited_by_percentile_year":null,"biblio":{"volume":null,"issue":null,"first_page":"23870","last_page":"23890"},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T10181","display_name":"Natural Language Processing Techniques","score":0.9092000126838684,"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/T10181","display_name":"Natural Language Processing Techniques","score":0.9092000126838684,"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/computer-science","display_name":"Computer science","score":0.7131130695343018},{"id":"https://openalex.org/keywords/inference","display_name":"Inference","score":0.6996122598648071},{"id":"https://openalex.org/keywords/artificial-intelligence","display_name":"Artificial intelligence","score":0.37187302112579346},{"id":"https://openalex.org/keywords/natural-language-processing","display_name":"Natural language processing","score":0.3631948232650757}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.7131130695343018},{"id":"https://openalex.org/C2776214188","wikidata":"https://www.wikidata.org/wiki/Q408386","display_name":"Inference","level":2,"score":0.6996122598648071},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.37187302112579346},{"id":"https://openalex.org/C204321447","wikidata":"https://www.wikidata.org/wiki/Q30642","display_name":"Natural language processing","level":1,"score":0.3631948232650757}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.18653/v1/2025.findings-acl.1223","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2025.findings-acl.1223","pdf_url":"https://aclanthology.org/2025.findings-acl.1223.pdf","source":null,"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Findings of the Association for Computational Linguistics: ACL 2025","raw_type":"proceedings-article"}],"best_oa_location":{"id":"doi:10.18653/v1/2025.findings-acl.1223","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2025.findings-acl.1223","pdf_url":"https://aclanthology.org/2025.findings-acl.1223.pdf","source":null,"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Findings of the Association for Computational Linguistics: ACL 2025","raw_type":"proceedings-article"},"sustainable_development_goals":[],"awards":[],"funders":[],"has_content":{"grobid_xml":true,"pdf":true},"content_urls":{"pdf":"https://content.openalex.org/works/W4412944721.pdf","grobid_xml":"https://content.openalex.org/works/W4412944721.grobid-xml"},"referenced_works_count":0,"referenced_works":[],"related_works":["https://openalex.org/W4391375266","https://openalex.org/W2899084033","https://openalex.org/W2748952813","https://openalex.org/W2390279801","https://openalex.org/W4391913857","https://openalex.org/W2358668433","https://openalex.org/W4396701345","https://openalex.org/W2376932109","https://openalex.org/W2001405890","https://openalex.org/W3204019825"],"abstract_inverted_index":{"Large":[0],"language":[1],"models":[2,27],"(LLMs)":[3],"have":[4,51],"demonstrated":[5],"impressive":[6],"performance":[7],"across":[8],"a":[9,75,118],"wide":[10],"range":[11],"of":[12],"tasks,":[13],"including":[14,144],"open-ended":[15],"dialogue,":[16],"driving":[17],"advancements":[18],"in":[19,43,110,145],"virtual":[20],"assistants":[21],"and":[22,39,159],"other":[23],"interactive":[24],"systems.However,":[25],"these":[26],"often":[28],"generate":[29],"outputs":[30],"misaligned":[31],"with":[32],"human":[33],"values,":[34],"such":[35],"as":[36,117,122],"ethical":[37],"norms":[38],"safety":[40,129,155],"constraints,":[41],"resulting":[42],"potentially":[44],"harmful":[45],"or":[46,65,121,134],"inappropriate":[47],"responses.While":[48],"several":[49],"techniques":[50],"been":[52],"proposed":[53],"to":[54,88,102,106],"address":[55],"this":[56,70],"problem,":[57],"they":[58],"typically":[59],"involve":[60],"computationally":[61],"intensive":[62],"training":[63],"procedures":[64],"introduce":[66],"substantial":[67],"inference-time":[68],"latency.In":[69],"paper,":[71],"we":[72,149],"present":[73],"DIESEL,":[74],"lightweight":[76],"inferenceguidance":[77],"technique":[78],"that":[79,151],"can":[80,114],"be":[81],"seamlessly":[82],"integrated":[83],"into":[84],"any":[85],"autoregressive":[86],"LLM":[87],"semantically":[89],"filter":[90],"undesirable":[91],"content":[92],"during":[93],"generation.DIESEL":[94],"guides":[95],"generation":[96],"by":[97],"reranking":[98],"token":[99],"candidates":[100],"according":[101],"their":[103],"semantic":[104],"similarity":[105],"predefined":[107],"negative":[108],"concepts":[109],"the":[111],"latent":[112],"space.It":[113],"serve":[115],"either":[116],"standalone":[119],"safeguard":[120],"an":[123],"auxiliary":[124],"defense":[125],"layer,":[126],"enhancing":[127],"response":[128,161],"without":[130],"requiring":[131],"model":[132],"fine-tuning":[133],"additional":[135],"data.We":[136],"demonstrate":[137],"DIESEL's":[138],"effectiveness":[139],"on":[140],"state-of-the-art":[141],"conversational":[142],"models,":[143],"adversarial":[146],"jailbreak":[147],"scenarios.Furthermore,":[148],"show":[150],"DIESEL":[152],"generalizes":[153],"beyond":[154],"applications,":[156],"enabling":[157],"flexible":[158],"domain-specific":[160],"filtering.":[162]},"counts_by_year":[],"updated_date":"2026-07-15T18:14:33.161393","created_date":"2025-10-10T00:00:00"}