{"id":"https://openalex.org/W4412888392","doi":"https://doi.org/10.18653/v1/2025.findings-acl.394","title":"Hierarchical Safety Realignment: Lightweight Restoration of Safety in Pruned Large Vision-Language Models","display_name":"Hierarchical Safety Realignment: Lightweight Restoration of Safety in Pruned Large Vision-Language Models","publication_year":2025,"publication_date":"2025-01-01","ids":{"openalex":"https://openalex.org/W4412888392","doi":"https://doi.org/10.18653/v1/2025.findings-acl.394"},"language":"en","primary_location":{"id":"doi:10.18653/v1/2025.findings-acl.394","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2025.findings-acl.394","pdf_url":"https://aclanthology.org/2025.findings-acl.394.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.394.pdf","any_repository_has_fulltext":null},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5100638570","display_name":"Yue Li","orcid":"https://orcid.org/0000-0001-7836-4144"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Yue Li","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5046852308","display_name":"Xin Yi","orcid":"https://orcid.org/0000-0002-4726-5765"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Xin Yi","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5022536070","display_name":"Dongsheng Shi","orcid":"https://orcid.org/0000-0002-5701-9795"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Dongsheng Shi","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5085818578","display_name":"Gerard de Melo","orcid":"https://orcid.org/0000-0002-2930-2059"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Gerard De Melo","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5042651811","display_name":"Xiaoling Wang","orcid":"https://orcid.org/0000-0002-5396-2284"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Xiaoling Wang","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5100349011","display_name":"Linlin Wang","orcid":"https://orcid.org/0000-0003-3678-5922"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Linlin Wang","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":1,"citation_normalized_percentile":null,"cited_by_percentile_year":null,"biblio":{"volume":null,"issue":null,"first_page":"7600","last_page":"7612"},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T10215","display_name":"Semantic Web and Ontologies","score":0.8258000016212463,"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/T10215","display_name":"Semantic Web and Ontologies","score":0.8258000016212463,"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":0.8065999746322632,"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/T10036","display_name":"Advanced Neural Network Applications","score":0.8047999739646912,"subfield":{"id":"https://openalex.org/subfields/1707","display_name":"Computer Vision and Pattern Recognition"},"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":0.680084228515625},{"id":"https://openalex.org/keywords/artificial-intelligence","display_name":"Artificial intelligence","score":0.33284980058670044}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.680084228515625},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.33284980058670044}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.18653/v1/2025.findings-acl.394","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2025.findings-acl.394","pdf_url":"https://aclanthology.org/2025.findings-acl.394.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.394","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2025.findings-acl.394","pdf_url":"https://aclanthology.org/2025.findings-acl.394.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":[{"display_name":"Sustainable cities and communities","id":"https://metadata.un.org/sdg/11","score":0.5699999928474426}],"awards":[],"funders":[],"has_content":{"grobid_xml":true,"pdf":true},"content_urls":{"pdf":"https://content.openalex.org/works/W4412888392.pdf","grobid_xml":"https://content.openalex.org/works/W4412888392.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/W4396696052"],"abstract_inverted_index":{"With":[0],"the":[1,57,66,91,98,103,121,128],"growing":[2],"size":[3],"of":[4,59,93,123],"Large":[5],"Vision-Language":[6],"Models":[7],"(LVLMs),":[8],"network":[9],"pruning":[10,29,112],"techniques":[11],"designed":[12],"to":[13,63,102],"compress":[14],"these":[15,77],"models":[16,110],"for":[17],"deployment":[18],"in":[19,32,35,85,118,136],"resource-constrained":[20],"environments":[21],"have":[22],"attracted":[23],"significant":[24],"attention.However,":[25],"we":[26,41],"observe":[27],"that":[28,80],"frequently":[30],"results":[31],"a":[33,43,82],"degradation":[34],"safety":[36,92,119,135],"performance.To":[37,120],"address":[38],"this":[39,126],"issue,":[40],"propose":[42],"novel":[44],"and":[45,70,111],"lightweight":[46],"approach,":[47],"named":[48],"Hierarchical":[49],"Safety":[50],"Realignment":[51],"(HSR).HSR":[52],"operates":[53],"by":[54],"first":[55,129],"quantifying":[56],"contribution":[58],"each":[60],"attention":[61,78,99],"head":[62,100],"safety,":[64],"identifying":[65],"most":[67],"critical":[68],"ones,":[69],"then":[71],"selectively":[72],"restoring":[73,134],"neurons":[74],"directly":[75],"within":[76],"heads":[79],"play":[81],"pivotal":[83],"role":[84],"maintaining":[86],"safety.This":[87],"process":[88],"hierarchically":[89],"realigns":[90],"pruned":[94],"LVLMs,":[95],"progressing":[96],"from":[97],"level":[101],"neuron":[104],"level.We":[105],"validate":[106],"HSR":[107],"across":[108],"various":[109],"strategies,":[113],"consistently":[114],"achieving":[115],"notable":[116],"improvements":[117],"best":[122],"our":[124],"knowledge,":[125],"is":[127],"work":[130],"explicitly":[131],"focused":[132],"on":[133],"LVLMs":[137],"post-pruning.The":[138],"code":[139],"will":[140],"be":[141],"available":[142],"at":[143],"https://github.com/TheShineyue/HSR.":[144]},"counts_by_year":[{"year":2026,"cited_by_count":1}],"updated_date":"2026-07-14T23:27:15.235271","created_date":"2025-10-10T00:00:00"}