[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"detail-sidebar-cat-0-en-105":3,"doc-seo-273476-105":59,"doc-detail-273476-en":131},{"code":4,"msg":5,"data":6},0,"success",[7,13,18,23,28,33,38,43,48,51,55],{"id":8,"doc_module":4,"doc_module_name":9,"category_name":10,"show_sort_weight":11,"slug":12},1,"Document","Story & Novel",90,"story-novel",{"id":14,"doc_module":4,"doc_module_name":9,"category_name":15,"show_sort_weight":16,"slug":17},2,"Literature",80,"literature",{"id":19,"doc_module":4,"doc_module_name":9,"category_name":20,"show_sort_weight":21,"slug":22},4,"Exam",70,"exam",{"id":24,"doc_module":4,"doc_module_name":9,"category_name":25,"show_sort_weight":26,"slug":27},5,"Comic",60,"comic",{"id":29,"doc_module":4,"doc_module_name":9,"category_name":30,"show_sort_weight":31,"slug":32},6,"Technology",50,"technology",{"id":34,"doc_module":4,"doc_module_name":9,"category_name":35,"show_sort_weight":36,"slug":37},7,"Healthcare",40,"healthcare",{"id":39,"doc_module":4,"doc_module_name":9,"category_name":40,"show_sort_weight":41,"slug":42},8,"Research & Report",30,"research-report",{"id":44,"doc_module":4,"doc_module_name":9,"category_name":45,"show_sort_weight":46,"slug":47},9,"Religion & Spirituality",20,"religion-spirituality",{"id":46,"doc_module":4,"doc_module_name":9,"category_name":49,"show_sort_weight":46,"slug":50},"World Cup","world-cup",{"id":52,"doc_module":4,"doc_module_name":9,"category_name":53,"show_sort_weight":52,"slug":54},10,"Lifestyle","lifestyle",{"id":56,"doc_module":4,"doc_module_name":9,"category_name":57,"show_sort_weight":24,"slug":58},19,"General","general",{"code":4,"msg":60,"data":61},"ok",{"site_id":62,"language":63,"slug":64,"title":65,"keywords":66,"description":67,"schema_data":68,"social_meta":124,"head_meta":126,"extra_data":128,"updated_unix":130},105,"en","babelsteering-multilingual-safety-alignment-via-english-steering-vectors-abstract","BabelSteering - Multilingual Safety Alignment via English Steering Vectors - Abstract","","Large language models are deployed worldwide in high-stakes settings, yet most safety alignment work targets English, leaving weaker safeguards for users in other languages. This paper examines whether safety signals learned in a high-resource language can generalize. BabelSteering is an activation steering method using English-derived refusal directions as a lightweight inference-time intervention. Evaluation across eight languages measures refusal of harmful requests, over-refusal, and task utility, improving harmful refusal with marginal utility loss and increasing pseudo-harmful refusals.",{"@graph":69,"@context":123},[70,84,106],{"@type":71,"itemListElement":72},"BreadcrumbList",[73,77,79,82],{"item":74,"name":75,"@type":76,"position":8},"https://docshare.wps.com","Home","ListItem",{"item":78,"name":9,"@type":76,"position":14},"https://docshare.wps.com/document/",{"item":80,"name":40,"@type":76,"position":81},"https://docshare.wps.com/document/research-report/",3,{"item":83,"name":65,"@type":76,"position":19},"https://docshare.wps.com/document/babelsteering-multilingual-safety-alignment-via-english-steering-vectors-abstract/273476/",{"url":83,"name":65,"@type":85,"image":86,"author":91,"headline":65,"publisher":94,"fileFormat":97,"inLanguage":63,"description":67,"dateModified":98,"datePublished":99,"encodingFormat":97,"isAccessibleForFree":100,"interactionStatistic":101},"DigitalDocument",{"url":87,"@type":88,"width":89,"height":90},"https://docshare.wps.com/thumbnails/babelsteering-multilingual-safety-alignment-via-english-steering-vectors-abstract/273476.png","ImageObject",300,407,{"name":92,"@type":93},"Emma Wilson","Person",{"url":74,"name":95,"@type":96},"DocShare","Organization","application/pdf","2026-11-04","2026-09-15",true,{"@type":102,"interactionType":103,"userInteractionCount":105},"InteractionCounter",{"@type":104},"ViewAction",11,{"@type":107,"mainEntity":108},"FAQPage",[109,115,119],{"name":110,"@type":111,"acceptedAnswer":112},"What problem does BabelSteering address in multilingual settings?","Question",{"text":113,"@type":114},"Safety research and alignment are concentrated in English, so users of other languages may face weaker safeguards despite using the same LLMs.","Answer",{"name":116,"@type":111,"acceptedAnswer":117},"How does BabelSteering work at inference time?",{"text":118,"@type":114},"It uses activation steering with refusal directions derived from English safety supervision, acting as a lightweight inference-time intervention to generalize across languages.",{"name":120,"@type":111,"acceptedAnswer":121},"What metrics are used to evaluate multilingual safety and usability?",{"text":122,"@type":114},"The evaluation jointly measures refusal of harmful requests, over-refusal, and general task utility across eight languages.","https://schema.org",{"og:url":83,"og:type":125,"og:title":65,"og:site_name":95,"og:description":67},"article",{"robots":127,"canonical":83},"index,follow",{"doc_id":129,"site_id":62},273476,1789471617,{"code":4,"msg":5,"data":132},{"doc_id":129,"user_id":133,"nickname":92,"user_avatar":134,"doc_module":4,"category_id":39,"category_name":40,"doc_title":65,"doc_description":67,"doc_content":135,"file_id":136,"file_url":137,"file_type":138,"file_size":139,"view_count":105,"is_deleted":4,"is_public":8,"is_downloadable":8,"audit_status":8,"page_count":140,"language":141,"language_code":63,"site_id":62,"html_lang":63,"table_of_contents":142,"faqs":143,"seo_title":144,"seo_description":67,"update_tm":130,"read_time":145,"preview_page_count":4},3848291630094,"https://eur-avatar.wpscdn.com/davatar_085a072bc5b1113ac321206ff7593b45","BabelSteering: Multilingual Safety Alignment via English Steering Vectors  \nEmma V. Stein* 1 Dominik Meier* 1 ,2 Terry Ruas 1 Jan Philip Wahle†1 Bela Gipp†1  \n1 University of Göttingen, Germany  \n2 German State Police NRW, Germany  \n[emma.stein@stud.uni-goettingen.de](emma.stein@stud.uni-goettingen.de) , [meier@gipplab.org](meier@gipplab.org)  \narXiv :2608 . 16577v1 [ cs .CL] 17 Aug 2026  \nAbstract  \nLarge language models (LLMs) are deployed globally in high-stakes settings, yet most safety research and alignment efforts remain concentrated on English. Thus, users interacting with LLMs in other languages may encounter weaker safeguards despite relying on the same systems for similarly sensitive tasks. In this work, we investigate whether safety signals learned from a high-resource language, like English, can improve multilingual safety. We propose BabelSteering, an activation steering method that acts as a lightweight inferencetime intervention, using refusal directions derived from English safety supervision to generalize across languages. Our evaluation includes eight languages and jointly measures refusal of harmful requests, over-refusal, and general task utility. The results show that BabelSteering increases the refusal of harmful requests across languages, with only a marginal to no reduction in task utility but with some increase in refusal of pseudo-harmful prompts. For example, for Gemma 7B, we see an average increase in the refusal of harmful prompts across languages of 11 percentage points (pp), with individual languages like Bengali seeing an increase of 17 pp, with no loss of utility on Global MMLU, while pseudo-harmful refusals increase by 13 pp on average. We also introduce a multilingual translation-and-evaluation pipeline to facilitate future work on cross-lingual safety interventions. Overall, our findings suggest that activation steering may provide a practical, lowcost mechanism for extending English-derived safety signals to other languages.  \nWarning: this paper contains examples with unsafe content.  \n1 The Multilingual Safety Gap  \nLarge language models (LLMs) are deployed in highstakes contexts worldwide, including health information, education, legal assistance, and public services (Jin et al., 2024 ; Cascella et al., 2023 ; Dahl et al., 2024 ;  \n*Equal contribution.†Jointly supervised this work.  \nJonnala et al., 2024 ; Meier et al., 2025) . As these systems reach a global user base of billions of people, their safety becomes important across many different languages. Nevertheless, safety research remains heavily concentrated on English (Yong et al., 2025) . The second most studied language, Chinese, has around 10 times fewer safety research papers than English, and, for example, on ChatbotArena, only 5 of 20 models that provide a system report have wide multilingual support, report their multi-lingual safety alignment training, and red-teaming efforts (Yong et al., 2025) . This imbalance creates a structural problem in which users interacting in non-English languages may encounter weaker safeguards, despite relying on the same systems for similarly sensitive tasks.  \nOne driver of this imbalance is the cost structure of current safety pipelines. Effective safety alignment often relies on carefully annotated training data (Ouyang et al., 2022 ; Dai et al., 2024) including high-quality user feedback on the safety of responses. This is already expensive for English and becomes hard to scale across many languages, particularly low-resource languages where access to first-language speakers is limited. One recent survey (Röttger et al., 2025) found that roughly 80% of open safety datasets are limited to English, with no non-English datasets specifically designed for model training. In practice, this slows the deployment of robust safeguards in underrepresented languages and makes continuous updates difficult as models and threat surfaces evolve. Furthermore, there is another barrier to training a model on","cbCaijmZ9ceNnNoq","https://ap.wps.com/l/cbCaijmZ9ceNnNoq","pdf",1016445,14,"English","# The Multilingual Safety Gap\n## Motivation: safety pipelines and data imbalance\n## Prior work and safety-geometry transfer\n# BabelSteering Approach\n## Activation steering with English refusal directions\n## Translation-and-evaluation pipeline\n# Evaluation\n## Metrics: harmful refusal, over-refusal, utility\n## Results across eight languages\n# Related Work Comparison\n## Table of multilingual safety-steering methods","[{\"question\":\"What problem does BabelSteering address in multilingual settings?\",\"answer\":\"Safety research and alignment are concentrated in English, so users of other languages may face weaker safeguards despite using the same LLMs.\"},{\"question\":\"How does BabelSteering work at inference time?\",\"answer\":\"It uses activation steering with refusal directions derived from English safety supervision, acting as a lightweight inference-time intervention to generalize across languages.\"},{\"question\":\"What metrics are used to evaluate multilingual safety and usability?\",\"answer\":\"The evaluation jointly measures refusal of harmful requests, over-refusal, and general task utility across eight languages.\"}]","BabelSteering - Multilingual Safety Alignment via English Steering Vectors - Abstract | PDF",35]