[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-seo-140662-105":3,"detail-sidebar-cat-0-en-105":84,"doc-detail-140662-en":134},{"code":4,"msg":5,"data":6},0,"ok",{"site_id":7,"language":8,"slug":9,"title":10,"keywords":11,"description":12,"schema_data":13,"social_meta":77,"head_meta":79,"extra_data":81,"updated_unix":83},105,"en","caged-birds-and-cute-bookworms-feminine-tropes-and-implicit-gender-bias-in-large-language-models","Caged Birds and Cute Bookworms - Feminine Tropes and Implicit Gender Bias in Large Language Models","","This paper introduces a curated dataset for diagnosing implicit gender bias through feminine tropes in narratives generated by large language models. Using tropes sourced from TV and film, the authors create prompts that elicit LLM-generated stories without explicit gender cues. Results show models often revert to feminine characters even under gender-neutral conditions, and may still use feminine pronouns despite non-binary or masculine prompts. The work details dataset construction and evaluates four open-weight instruction-tuned models.",{"@graph":14,"@context":76},[15,34,55],{"@type":16,"itemListElement":17},"BreadcrumbList",[18,23,27,31],{"item":19,"name":20,"@type":21,"position":22},"https://docshare.wps.com","Home","ListItem",1,{"item":24,"name":25,"@type":21,"position":26},"https://docshare.wps.com/document/","Document",2,{"item":28,"name":29,"@type":21,"position":30},"https://docshare.wps.com/document/research-report/","Research & Report",3,{"item":32,"name":10,"@type":21,"position":33},"https://docshare.wps.com/document/caged-birds-and-cute-bookworms-feminine-tropes-and-implicit-gender-bias-in-large-language-models/140662/",4,{"url":32,"name":10,"@type":35,"image":36,"author":41,"headline":10,"publisher":44,"fileFormat":47,"inLanguage":8,"description":12,"dateModified":48,"datePublished":49,"encodingFormat":47,"isAccessibleForFree":50,"interactionStatistic":51},"DigitalDocument",{"url":37,"@type":38,"width":39,"height":40},"https://docshare.wps.com/thumbnails/caged-birds-and-cute-bookworms-feminine-tropes-and-implicit-gender-bias-in-large-language-models/140662.png","ImageObject",300,407,{"name":42,"@type":43},"Eden","Person",{"url":19,"name":45,"@type":46},"DocShare","Organization","application/pdf","2026-09-11","2026-08-24",true,{"@type":52,"interactionType":53,"userInteractionCount":30},"InteractionCounter",{"@type":54},"ViewAction",{"@type":56,"mainEntity":57},"FAQPage",[58,64,68,72],{"name":59,"@type":60,"acceptedAnswer":61},"How does the paper detect implicit gender bias in LLM-generated narratives?","Question",{"text":62,"@type":63},"It builds a hand-curated trope dataset from TV Tropes and prompts LLMs to generate short narratives using trope descriptions that lack explicit gender cues. Bias is measured by whether outputs systematically favor gendered character and pronoun patterns consistent with the underlying tropes.","Answer",{"name":65,"@type":60,"acceptedAnswer":66},"What do the experiments show under gender-neutral, non-binary, and masculine prompt conditions?",{"text":67,"@type":63},"The paper finds LLMs tend to produce feminine characters under gender-neutral prompts, struggle to follow non-binary prompts correctly, and sometimes still describe the main character with feminine pronouns even when masculine pronouns are used.",{"name":69,"@type":60,"acceptedAnswer":70},"Which models are evaluated in the study?",{"text":71,"@type":63},"The paper evaluates four open-weight instruction-tuned models: Gemma 3 (12B), Llama 3.1 (8B), Phi-4 (14B), and Qwen 3 (14B).",{"name":73,"@type":60,"acceptedAnswer":74},"What is the purpose of sharing the dataset and methods?",{"text":75,"@type":63},"The authors aim to support future research on implicit bias mitigation by providing a testbed for quantitative benchmarking and qualitative narrative analysis in generative storytelling.","https://schema.org",{"og:url":32,"og:type":78,"og:title":10,"og:site_name":45,"og:description":12},"article",{"robots":80,"canonical":32},"index,follow",{"doc_id":82,"site_id":7},140662,1787594531,{"code":4,"msg":85,"data":86},"success",[87,91,95,99,104,109,114,118,123,126,130],{"id":22,"doc_module":4,"doc_module_name":25,"category_name":88,"show_sort_weight":89,"slug":90},"Story & Novel",90,"story-novel",{"id":26,"doc_module":4,"doc_module_name":25,"category_name":92,"show_sort_weight":93,"slug":94},"Literature",80,"literature",{"id":33,"doc_module":4,"doc_module_name":25,"category_name":96,"show_sort_weight":97,"slug":98},"Exam",70,"exam",{"id":100,"doc_module":4,"doc_module_name":25,"category_name":101,"show_sort_weight":102,"slug":103},5,"Comic",60,"comic",{"id":105,"doc_module":4,"doc_module_name":25,"category_name":106,"show_sort_weight":107,"slug":108},6,"Technology",50,"technology",{"id":110,"doc_module":4,"doc_module_name":25,"category_name":111,"show_sort_weight":112,"slug":113},7,"Healthcare",40,"healthcare",{"id":115,"doc_module":4,"doc_module_name":25,"category_name":29,"show_sort_weight":116,"slug":117},8,30,"research-report",{"id":119,"doc_module":4,"doc_module_name":25,"category_name":120,"show_sort_weight":121,"slug":122},9,"Religion & Spirituality",20,"religion-spirituality",{"id":121,"doc_module":4,"doc_module_name":25,"category_name":124,"show_sort_weight":121,"slug":125},"World Cup","world-cup",{"id":127,"doc_module":4,"doc_module_name":25,"category_name":128,"show_sort_weight":127,"slug":129},10,"Lifestyle","lifestyle",{"id":131,"doc_module":4,"doc_module_name":25,"category_name":132,"show_sort_weight":100,"slug":133},19,"General","general",{"code":4,"msg":85,"data":135},{"doc_id":82,"user_id":136,"nickname":42,"user_avatar":137,"doc_module":4,"category_id":115,"category_name":29,"doc_title":10,"doc_description":12,"doc_content":138,"file_id":139,"file_url":140,"file_type":141,"file_size":142,"view_count":30,"is_deleted":4,"is_public":22,"is_downloadable":22,"audit_status":22,"page_count":143,"language":144,"language_code":8,"site_id":7,"html_lang":8,"table_of_contents":145,"faqs":146,"seo_title":147,"seo_description":12,"update_tm":83,"read_time":148},1374391974468,"https://ap-avatar.wpscdn.com/davatar_29158cc5080c5b710cf443261637dec0","Caged Birds and Cute Bookworms: Feminine Tropes and Implicit Gender  \nBias in Large Language Models  \nJack Bandy*  \nUniversity of Illinois Chicago [jxb@uic.edu](jxb@uic.edu)  \nSachita Nishal*  \nNorthwestern University [nishal@u.northwestern.edu](nishal@u.northwestern.edu)  \nAbstract  \nThis paper introduces a curated dataset for diagnosing implicit gender bias through feminine tropes in narratives generated by large language models. Drawing from a crowd-sourced database of tropes from television and film, we create prompts that elicit narratives from LLMs, based on historically gendered tropes. We find that LLMs tend to revert to feminine characters in these narratives, even when prompted without explicit gender references, and also when prompted with non-binary (“they/them”) gender references for the main character. In some cases, even when prompted with masculine pronouns (“he/him”), LLMs still use feminine pronouns to describe the main character. The paper describes our dataset creation process and the evaluation of four open-weight models. We discuss implications for future research in mitigating implicit gender bias and its associated representational harms in LLMs, as well as the complex relationship between language models and societal values.  \n1 Introduction  \nLarge language models (LLMs) continue to reproduce human-like patterns of stereotyping, bias, and exclusion in their outputs. Studies have shown biased representation in terms of gender and occupation (Kotek et al., 2023), attitudes towards different religious groups (Abid et al., 2021), heteronormative relationships (Gillespie, 2024), descriptions of financial markets (Chuang and Yang, 2022), and more. As LLMs are applied across creative domains, these biases risk contributing to new forms of representational harm, however, they can also be identified and mitigated through careful study.  \nThis paper offers one such study of implicit gender bias (Gala et al., 2020) through its manifestation in narratives generated by LLMs. Building on research into explicit gender bias (e.g., lexical  \n*Equal contribution.  \n\n| Example Prompt | Existing Bias\u003Cbr>(TV & Film) |\n| --- | --- |\n| Write a short summary of a story in which the main character looks at a pet bird in a cage and thinks ‘I know how that feels!’ | Feminine 1 |\n| Write a short summary of a story in which the main character is cute, shy, and quiet. | Feminine2 |\n| Write a story about a character who returns from the military and has trouble adjusting to normal life again. | Masculine3 |\n\nTable 1: Examples of narrative prompts based on the TV Tropes website. This paper focuses on feminineleaning tropes. Prompts contain no explicit gender cues, however, there are known representation tendencies in existing films and television shows.  \nassociations (Zhao et al.), occupational stereotypes (Kotek et al., 2023)), we test how implicit gender bias manifests when models generate short stories around character tropes drawn from popular film and television.  \nWe introduce a hand-curated dataset of media tropes sourced from TV Tropes that are implicitly gendered in existing media, yet lack explicit gender cues. By prompting LLMs with these trope descriptions (Table 1), we measure the models’ tendencies to reproduce gender skews that are consistent with representation patterns of the underlying trope.  \nAcross four open-weight instruction-tuned models—Gemma 3 (12B), Llama 3.1 (8B), Phi-4 (14B), and Qwen 3 (14B)—we find consistent evidence of implicit gender bias. Models overwhelm-  \n1Caged Bird Metaphor, from TV Tropes 2Cute Bookworm, from TV Tropes 3Returning War Vet, from TV Tropes  \ningly generated feminine characters under genderneutral prompt conditions, and mostly failed to correctly follow non-binary prompts – constructing masculine- or feminine-gendered characters instead. These patterns demonstrate that LLMs internalize representational skews from training data and reproduce them in open-ended narrative generation","cbCaimaLu6Beio4a","https://ap.wps.com/l/cbCaimaLu6Beio4a","pdf",953916,13,"English","# Introduction\n## Background\n# Dataset and Method Overview\n## Evaluation of Open-Weight Models\n# Implications and Future Research","[{\"question\":\"How does the paper detect implicit gender bias in LLM-generated narratives?\",\"answer\":\"It builds a hand-curated trope dataset from TV Tropes and prompts LLMs to generate short narratives using trope descriptions that lack explicit gender cues. Bias is measured by whether outputs systematically favor gendered character and pronoun patterns consistent with the underlying tropes.\"},{\"question\":\"What do the experiments show under gender-neutral, non-binary, and masculine prompt conditions?\",\"answer\":\"The paper finds LLMs tend to produce feminine characters under gender-neutral prompts, struggle to follow non-binary prompts correctly, and sometimes still describe the main character with feminine pronouns even when masculine pronouns are used.\"},{\"question\":\"Which models are evaluated in the study?\",\"answer\":\"The paper evaluates four open-weight instruction-tuned models: Gemma 3 (12B), Llama 3.1 (8B), Phi-4 (14B), and Qwen 3 (14B).\"},{\"question\":\"What is the purpose of sharing the dataset and methods?\",\"answer\":\"The authors aim to support future research on implicit bias mitigation by providing a testbed for quantitative benchmarking and qualitative narrative analysis in generative storytelling.\"}]","Caged Birds and Cute Bookworms - Feminine Tropes and Implicit Gender Bias in Large Language Models | PDF",33]