[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"detail-sidebar-cat-1-en-105":3,"doc-seo-204916-105":53,"doc-detail-204916-en":127},{"code":4,"msg":5,"data":6},0,"success",[7,14,19,24,29,34,39,44,49],{"id":8,"doc_module":9,"doc_module_name":10,"category_name":11,"show_sort_weight":12,"slug":13},11,1,"Template","Presentations",90,"presentations",{"id":15,"doc_module":9,"doc_module_name":10,"category_name":16,"show_sort_weight":17,"slug":18},12,"Resumes",80,"resumes",{"id":20,"doc_module":9,"doc_module_name":10,"category_name":21,"show_sort_weight":22,"slug":23},14,"Invoices",70,"invoices",{"id":25,"doc_module":9,"doc_module_name":10,"category_name":26,"show_sort_weight":27,"slug":28},15,"Posters",60,"posters",{"id":30,"doc_module":9,"doc_module_name":10,"category_name":31,"show_sort_weight":32,"slug":33},16,"Social Media",50,"social-media",{"id":35,"doc_module":9,"doc_module_name":10,"category_name":36,"show_sort_weight":37,"slug":38},17,"Forms",40,"forms",{"id":40,"doc_module":9,"doc_module_name":10,"category_name":41,"show_sort_weight":42,"slug":43},18,"Letters",30,"letters",{"id":45,"doc_module":9,"doc_module_name":10,"category_name":46,"show_sort_weight":47,"slug":48},21,"Paper Templates",5,"papers-templates",{"id":50,"doc_module":9,"doc_module_name":10,"category_name":51,"show_sort_weight":4,"slug":52},158,"General","general-158",{"code":4,"msg":54,"data":55},"ok",{"site_id":56,"language":57,"slug":58,"title":59,"keywords":60,"description":61,"schema_data":62,"social_meta":120,"head_meta":122,"extra_data":124,"updated_unix":126},105,"en","efficient-model-editing-for-safe-information-localization-and-stitching-nsf-workshop-on-safe-information-localization","Efficient Model Editing for Safe Information Localization and Stitching - NSF Workshop on Safe Information Localization","","Efficient Model Editing with task vectors enables multi-task adaptation by learning edits via addition and forgetting via negation, then merging task-specific knowledge through sparse masks. Experiments with RoBERTa-base across GLUE tasks show computational efficiency, reduced interference, and mechanistic interpretability via localized regions, especially in LayerNorm parameters. The approach further targets safety: cross-lingual fine-tuning attacks and stitching attacks reveal how localized “jailbroken” regions transfer across multilingual LLMs, explaining shared safety knowledge and vulnerabilities.",{"@graph":63,"@context":119},[64,80,102],{"@type":65,"itemListElement":66},"BreadcrumbList",[67,71,74,77],{"item":68,"name":69,"@type":70,"position":9},"https://docshare.wps.com","Home","ListItem",{"item":72,"name":10,"@type":70,"position":73},"https://docshare.wps.com/template/",2,{"item":75,"name":11,"@type":70,"position":76},"https://docshare.wps.com/template/presentations/",3,{"item":78,"name":59,"@type":70,"position":79},"https://docshare.wps.com/template/efficient-model-editing-for-safe-information-localization-and-stitching-nsf-workshop-on-safe-information-localization/204916/",4,{"url":78,"name":59,"@type":81,"image":82,"author":87,"headline":59,"publisher":90,"fileFormat":93,"inLanguage":57,"description":61,"dateModified":94,"datePublished":95,"encodingFormat":93,"isAccessibleForFree":96,"interactionStatistic":97},"DigitalDocument",{"url":83,"@type":84,"width":85,"height":86},"https://docshare.wps.com/thumbnails/efficient-model-editing-for-safe-information-localization-and-stitching-nsf-workshop-on-safe-information-localization/204916.png","ImageObject",442,249,{"name":88,"@type":89},"Pentious","Person",{"url":68,"name":91,"@type":92},"DocShare","Organization","application/pdf","2026-10-29","2026-09-05",true,{"@type":98,"interactionType":99,"userInteractionCount":101},"InteractionCounter",{"@type":100},"ViewAction",7,{"@type":103,"mainEntity":104},"FAQPage",[105,111,115],{"name":106,"@type":107,"acceptedAnswer":108},"What are task vectors and how are they used for model editing?","Question",{"text":109,"@type":110},"Task vectors represent the parameter change from fine-tuning; edits are combined with arithmetic operations such as addition and negation to learn new behaviors and remove old ones.","Answer",{"name":112,"@type":107,"acceptedAnswer":113},"What does Localize-and-Stitch do in the experiments?",{"text":114,"@type":110},"It uses sparse masks to localize model regions related to tasks, then stitches localized components to study effects on merged performance and interpretability.",{"name":116,"@type":107,"acceptedAnswer":117},"How do cross-lingual fine-tuning and stitching attacks relate to safety information localization?",{"text":118,"@type":110},"Adversarial fine-tuning in one language can disrupt multilingual safety alignment, while stitching localized jailbroken regions onto benign models breaks safety for unseen languages by exploiting shared safety-related regions.","https://schema.org",{"og:url":78,"og:type":121,"og:title":59,"og:site_name":91,"og:description":61},"article",{"robots":123,"canonical":78},"index,follow",{"doc_id":125,"site_id":56},204916,1791498352,{"code":4,"msg":5,"data":128},{"doc_id":125,"user_id":129,"nickname":88,"user_avatar":130,"doc_module":9,"category_id":8,"category_name":11,"doc_title":59,"doc_description":61,"doc_content":131,"file_id":132,"file_url":133,"file_type":134,"file_size":135,"view_count":101,"is_deleted":4,"is_public":9,"is_downloadable":9,"audit_status":9,"page_count":136,"language":137,"language_code":57,"site_id":56,"html_lang":57,"table_of_contents":138,"faqs":139,"seo_title":140,"seo_description":61,"update_tm":141,"read_time":76},1374404730887,"https://ap-avatar.wpscdn.com/davatar_6f874abed73319feea01a86fa6f0fab8","# Efficient Model Editing for Safe InformationLocalization and Stitching\n\nHan Zhao  \n02/26/2025  \n# NSF Workshop on Safe Al\n\nJoint work with Yifei He,Siqi Zeng,Yuzheng Hu,Rui Yang and Tong Zhang  \n工  \nhanzhao@illinois.edu  \nILLINOIS  \nDepartment of Computer ScienceUniversity of IIlinois Urbana-Champaign  \nComputer Science  \nTask vectors and model merging:  \nc)Learning via addition  \nb)Forgetting via negation  \nd)Task analogies  \na)Task vectors  \nTnew=TA+TB  \nTnew=TC+(TB—TA)  \nTnew=—T  \nExample:making a  \nExample:building a  \nExample:improving  \nlanguage model produce  \nT=θft-θpre  \nmulti-task model  \ndomain generalization  \nless toxic content  \nFine-tune k different tasks from a pre-trained model  \nto obtainfine-tuned model parameters θ;,Vi∈[k]  \n-Task vector  \n-Arithmetic operations on task vectors for multi-tasking capability:  \nfor some  \nFigure credit to Ilharco et al.,“Editing models with task arithmetic”  \n## Sparse task vectors for model merging?\n\n-Computationally more efficient  \n-Less task interference among task vectors,hence suboptimal  \n# multi-task performance of the merged model\n\n-Mechanistic interpretability:more transparent and explainable(?)  \n## Localize-and-Stitch\n\nExperiments  \n-Base models:RoBERTa-base  \n-Tasks:12 GLUE  \n-Sparsity:~1%-10%  \n1.0SST-2-1.000.190.280.070.040.050.010.020.020.030.030.03CR-0.191.000.210.100.040.050.010.010.010.030.040.04MR-0.280.211.000.120.040.060.010.010.010.030.040.040.8MPQA-0.070.100.121.000.040.060.000.010.010.020.040.06TREC-0.040.040.040.041.000.050.010.010.010.030.040.05  \n1.0SST-2-1.000.940.960.890.070.110.100.110.090.040.070.16CR-0.941.000.950.910.050.100.100.120.110.110.130.11MR-0.960.951.000.900.030.140.070.100.040.020.090.130.8MPQA-0.890.910.901.000.000.060.080.120.130.100.140.06TREC-0.070.050.030.001.000.140.080.100.090.000.160.00  \n0.6SUBJ-0.050.050.060.060.051.000.010.010.010.030.040.05QNLI-0.010.010.010.000.010.011.000.060.070.030.010.01  \n0.6SUBJ-0.110.100.140.060.141.000.060.130.070.050.150.03QNLI-0.100.100.070.080.080.061.000.310.410.580.530.61  \nSNLI-0.110.120.100.120.100.130.311.000.780.710.340.550.4MNLI-0.090.110.040.130.090.070.410.781.000.700.480.56RTE-0.040.110.020.100.000.050.580.710.701.000.540.64-0.2MRPC-0.070.130.090.140.160.150.530.340.480.541.000.51QQP-0.160.110.130.060.000.030.610.550.560.640.511.00  \nSNLI-0.020.010.010.010.010.010.061.000.100.050.010.010.4MNLI-0.020.010.010.010.010.010.070.101.000.040.010.01RTE-0.030.030.030.020.030.030.030.050.041.000.040.03  \n-0.2MRPC-0.030.040.040.040.040.040.010.010.010.041.000.06QQP-0.030.040.040.060.050.050.010.010.010.030.061.00意票  \n-0.0  \n-7NQ  \nSUB-  \nSNL-  \nI7NO  \nMRPC  \nQQP  \nRTE  \nMRPC  \nI7NW  \nRTE-  \nQQP  \nSNLI  \n(b)Cosine similarity of masked task vectors.  \n(a)Jaccard similarity of pairwise task masks.  \nLocalized modules:  \n(a)Distribution of localized regions in different network(b)Distribution of localized regions in different networklayers in the RoBERTa-base model.  \ncomponents in the RoBERTa-base model.  \n-The localized regions are predominantly found in the LayerNorm parameters  \n-Percentage =%of parameters in each module localized by the mask  \n# Safety Information Localization\n\nCross-lingual fine-tuning attacks:  \n-Fine-tuning on adversarial examples in one language breaksmultilingual safety alignment  \nMultiJail  \nAya Redteaming  \n\n| 0   | 6.1   | 0   | 9   | 3.4   |\n| --- | --- | --- | --- | --- |\n| 50.8   | 50.2   | 48.6   | 40   | 42.5  \u003Cbr>42.5   |\n| 45.4   | 47.9   | 44.1   | 45.1   |  |\n| 52.1   | 45.4   | 50.5   | 42.5   | 39.7  \u003Cbr>39   |\n| 36.2  \u003Cbr>35.2   |  | 34.3   | 47.3   |  |\n| 53.6     | 50.2     | 43.8     | 42.2   | 49.2   |\n\n\n| 7.6   | 2.2   | 5.5   | 1.7   | 8.9   |\n| --- | --- | --- | --- | --- |\n| 49.7   | 65.4   | 57.9   | 60.6   | 37.7   |\n| 52.9   | 67.5   | 62.5  \u003Cbr>62.3   |  | 39.4  \u003Cbr>39.7  \u003Cbr>38.6   |\n| 51.2   | 68.5   | 51.2   | 57.3   |  |\n| 38.3   | 72.2   | 53   | 45.5   |  |\n| 48.8   | 69.2   | 60.4    |  59.5   | 40.9   |\n\npre  \nen pre  \n60  \nen  \nQwen","cbCaipBFbRxEiePm","https://ap.wps.com/l/cbCaipBFbRxEiePm","pdf",8515768,9,"English","# Efficient Model Editing for Safe Information Localization and Stitching\n## Task vectors and model merging\n## Sparse task vectors for model merging?\n## Localize-and-Stitch\n## Safety Information Localization\n## Why do Task Vectors Work?\n## Ongoing projects","[{\"question\":\"What are task vectors and how are they used for model editing?\",\"answer\":\"Task vectors represent the parameter change from fine-tuning; edits are combined with arithmetic operations such as addition and negation to learn new behaviors and remove old ones.\"},{\"question\":\"What does Localize-and-Stitch do in the experiments?\",\"answer\":\"It uses sparse masks to localize model regions related to tasks, then stitches localized components to study effects on merged performance and interpretability.\"},{\"question\":\"How do cross-lingual fine-tuning and stitching attacks relate to safety information localization?\",\"answer\":\"Adversarial fine-tuning in one language can disrupt multilingual safety alignment, while stitching localized jailbroken regions onto benign models breaks safety for unseen languages by exploiting shared safety-related regions.\"}]","Efficient Model Editing for Safe Information Localization and Stitching - NSF Workshop on Safe Information Localization | PDF",1788576472]