[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"detail-sidebar-cat-0-en-105":3,"doc-seo-141387-105":59,"doc-detail-141387-en":122},{"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":115,"head_meta":117,"extra_data":119,"updated_unix":121},105,"en","biasdora-exploring-hidden-biased-associations-in-vision-language-models-abstract","BiasDora - Exploring Hidden Biased Associationsin Vision-Language Models - Abstract","","The work examines Vision-Language Models’ social biases beyond commonly documented stereotypes, which rely on narrow, pre-defined associations. It probes VLMs across nine bias dimensions using multiple input and output modalities, revealing hidden implicit associations. The study characterizes how biased associations differ in negativity, toxicity, and extremity, and identifies subtle as well as extreme biases not captured by prior methods, releasing the Dora dataset of retrieved associations.",{"@graph":69,"@context":114},[70,84,105],{"@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/biasdora-exploring-hidden-biased-associations-in-vision-language-models-abstract/141387/",{"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/biasdora-exploring-hidden-biased-associations-in-vision-language-models-abstract/141387.png","ImageObject",300,407,{"name":92,"@type":93},"Finn","Person",{"url":74,"name":95,"@type":96},"DocShare","Organization","application/pdf","2026-10-05","2026-08-25",true,{"@type":102,"interactionType":103,"userInteractionCount":39},"InteractionCounter",{"@type":104},"ViewAction",{"@type":106,"mainEntity":107},"FAQPage",[108],{"name":109,"@type":110,"acceptedAnswer":111},"What does the study contribute in addition to analysis results?","Question",{"text":112,"@type":113},"It publicly releases the Dataset of retrieved associations (Dora), containing the retrieved biased associations for further research and mitigation.","Answer","https://schema.org",{"og:url":83,"og:type":116,"og:title":65,"og:site_name":95,"og:description":67},"article",{"robots":118,"canonical":83},"index,follow",{"doc_id":120,"site_id":62},141387,1787655269,{"code":4,"msg":5,"data":123},{"doc_id":120,"user_id":124,"nickname":92,"user_avatar":125,"doc_module":4,"category_id":39,"category_name":40,"doc_title":65,"doc_description":67,"doc_content":126,"file_id":127,"file_url":128,"file_type":129,"file_size":130,"view_count":39,"is_deleted":4,"is_public":8,"is_downloadable":8,"audit_status":8,"page_count":131,"language":132,"language_code":63,"site_id":62,"html_lang":63,"table_of_contents":133,"faqs":134,"seo_title":135,"seo_description":67,"update_tm":121,"read_time":136},34359740700684,"https://ap-avatar.wpscdn.com/avatar/1f400023980c374ae676?_k=1777273430885731487","BiasDora: Exploring Hidden Biased Associationsin Vision-Language Models  \nNote: This paper contains examples of potentially offensive text and images generated by VLMs.  \nChahat Raj1 Anjishnu Mukherjee1 Aylin Caliskan2 Antonios Anastasopoulos1,3 Ziwei Zhu1  \n1 George Mason University, 2University of Washington  \n3Archimedes AI Research Unit, Athena RC, Greece  \n{craj,amukher6,antonis, [ziwei}@gmu. edu](ziwei}@gmu. edu) [aylin@uw. edu](aylin@uw. edu)  \nAbstract  \nExisting works examining Vision-Language Models (VLMs) for social biases predominantly focus on a limited set of documented bias associations, such as gender$profession  or  race$crime . This narrow scope often overlooks a vast range of unexamined implicit associations, restricting the identification and, hence, mitigation of such biases. We address this gap by probing VLMs to (1) uncover hidden, implicit associations across 9 bias dimensions. We systematically explore diverse input and output modalities and (2) demonstrate how biased associations vary in their negativity, toxicity, and extremity. Our work (3) identifies subtle and extreme biases that are typically not recognized by existing methodologies. We make the Dataset of retrieved associations,(Dora), publicly available.1  \n1 Introduction  \nDespite the transformative potential of VisionLanguage Models (VLMs) across many domains, mounting evidence underscored their risks to perpetuate and exacerbate social biases (Wan et al., 2024 ; Sathe et al., 2024), from reinforcing gender stereotypes by associating women with specific professions (Wan and Chang, 2024) to marginalizing minority communities by linking people of color with negative connotations (Ghosh and Caliskan, 2023) . Towards this, several bias evaluation methods have been designed (Caliskan et al., 2017 ; Nadeem et al., 2021a ; Howard et al., 2024 ; Smith et al., 2022 ; Hall et al., 2023) .  \nHowever, a critical limitation of existing evaluation methods is that they heavily rely on predefined associations like man$doctor and woman$nurse (Wan and Chang, 2024), remarkably narrowing their scope. The lists of associa-  \n1Data and code are available here [https://github](https://github) . com/chahatraj/BiasDora  \nFigure 1: VLMs reinforce biases that are different from the documented stereotypical associations.  \ntions2 in existing works represent just the tip of the iceberg in the vast spectrum of real-world biases. While most recent studies focus on evaluating occupational biases across different genders (Seshadriet al., 2023), Bansal et al. (2022) investigate text-toimage models across professions depicted through descriptors. Naik and Nushi (2023); Bianchi et al.(2023); Mandal et al. (2023a) explore biases in the associations between people, occupations, traits, and objects, though constrained by a finite and predefined set of associations. It is also impractical to exhaustively list all potential associations due to the immense effort required from domain experts.  \nMore importantly, the ultimate goal in assessing social biases in VLMs is to uncover all hidden biases within these models that can potentially harm individuals and society, not merely to confirm already known biases. Models may harbor biases that differ from those recognized by humans. There is an overlap between real-world biases and those inherent in VLMs (Figure 1), yet there is also a substantial portion of biases unique to VLMs that remain unexplored.  \n2The terms “biases” and “associations” are used interchangeably in this paper.  \n10439  \nFindings of the Association for Computational Linguistics: EMNLP 2024 , pages 10439–10455  \nNovember 12-16, 2024 ©2024 Association for Computational Linguistics  \nFigure 2: We probe VLMs in three modalities: T2T, T2I & I2T through word completion, image generation, and image description tasks. We calculate statistically significant association followed by identifying sentiment-negative and toxic association. We further evaluate bias levels of these associatio","cbCaik6sfQJf4Kb8","https://ap.wps.com/l/cbCaik6sfQJf4Kb8","pdf",8302559,17,"English","# Abstract\n# Introduction\n# VLM Probing","[{\"question\":\"What does the study contribute in addition to analysis results?\",\"answer\":\"It publicly releases the Dataset of retrieved associations (Dora), containing the retrieved biased associations for further research and mitigation.\"}]","BiasDora - Exploring Hidden Biased Associationsin Vision-Language Models - Abstract | PDF",43]