[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-seo-273355-105":3,"detail-sidebar-cat-0-en-105":81,"doc-detail-273355-en":130},{"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":74,"head_meta":76,"extra_data":78,"updated_unix":80},105,"en","cultrace-tracing-internal-cultural-reasoning-in-large-language-models-mechanistic-interpretability-framework","CulTrace - Tracing Internal Cultural Reasoning in Large Language Models - mechanistic interpretability framework","","CulTrace addresses the need to understand how large language models represent culture internally, beyond evaluating cultural awareness only through output analysis. Prior work can indicate when models answer incorrectly, but not why, because it ignores cultural information stored in model parameters. CulTrace is a mechanistic interpretability framework that probes layer-wise cultural knowledge processing and integration during cultural question answering, revealing staged reasoning and imbalanced culture resolution.",{"@graph":14,"@context":73},[15,34,56],{"@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/cultrace-tracing-internal-cultural-reasoning-in-large-language-models-mechanistic-interpretability-framework/273355/",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/cultrace-tracing-internal-cultural-reasoning-in-large-language-models-mechanistic-interpretability-framework/273355.png","ImageObject",300,407,{"name":42,"@type":43},"Jiven","Person",{"url":19,"name":45,"@type":46},"DocShare","Organization","application/pdf","2026-10-17","2026-09-15",true,{"@type":52,"interactionType":53,"userInteractionCount":55},"InteractionCounter",{"@type":54},"ViewAction",10,{"@type":57,"mainEntity":58},"FAQPage",[59,65,69],{"name":60,"@type":61,"acceptedAnswer":62},"What limitation in existing cultural alignment evaluations does CulTrace address?","Question",{"text":63,"@type":64},"Existing evaluations largely analyze model outputs, showing when responses are culturally inaccurate but offering limited insight into why. CulTrace focuses on how cultural information is represented and processed internally.","Answer",{"name":66,"@type":61,"acceptedAnswer":67},"How does CulTrace probe cultural reasoning inside an LLM?",{"text":68,"@type":64},"CulTrace trains layer-specific decoders to verbalize intermediate model representations. It generates chain-of-reasonings across layers to show when the model engages the domain, resolves cultural context, and selects an answer.",{"name":70,"@type":61,"acceptedAnswer":71},"What key pattern does CulTrace find about cultural reasoning across layers?",{"text":72,"@type":64},"It finds a consistent staged trajectory: models first engage the question’s domain, then resolve the relevant culture, and finally narrow in on an answer. Less-represented cultures tend to be resolved later and show more confusion.","https://schema.org",{"og:url":32,"og:type":75,"og:title":10,"og:site_name":45,"og:description":12},"article",{"robots":77,"canonical":32},"index,follow",{"doc_id":79,"site_id":7},273355,1789471027,{"code":4,"msg":82,"data":83},"success",[84,88,92,96,101,106,111,115,120,123,126],{"id":22,"doc_module":4,"doc_module_name":25,"category_name":85,"show_sort_weight":86,"slug":87},"Story & Novel",90,"story-novel",{"id":26,"doc_module":4,"doc_module_name":25,"category_name":89,"show_sort_weight":90,"slug":91},"Literature",80,"literature",{"id":33,"doc_module":4,"doc_module_name":25,"category_name":93,"show_sort_weight":94,"slug":95},"Exam",70,"exam",{"id":97,"doc_module":4,"doc_module_name":25,"category_name":98,"show_sort_weight":99,"slug":100},5,"Comic",60,"comic",{"id":102,"doc_module":4,"doc_module_name":25,"category_name":103,"show_sort_weight":104,"slug":105},6,"Technology",50,"technology",{"id":107,"doc_module":4,"doc_module_name":25,"category_name":108,"show_sort_weight":109,"slug":110},7,"Healthcare",40,"healthcare",{"id":112,"doc_module":4,"doc_module_name":25,"category_name":29,"show_sort_weight":113,"slug":114},8,30,"research-report",{"id":116,"doc_module":4,"doc_module_name":25,"category_name":117,"show_sort_weight":118,"slug":119},9,"Religion & Spirituality",20,"religion-spirituality",{"id":118,"doc_module":4,"doc_module_name":25,"category_name":121,"show_sort_weight":118,"slug":122},"World Cup","world-cup",{"id":55,"doc_module":4,"doc_module_name":25,"category_name":124,"show_sort_weight":55,"slug":125},"Lifestyle","lifestyle",{"id":127,"doc_module":4,"doc_module_name":25,"category_name":128,"show_sort_weight":97,"slug":129},19,"General","general",{"code":4,"msg":82,"data":131},{"doc_id":79,"user_id":132,"nickname":42,"user_avatar":133,"doc_module":4,"category_id":112,"category_name":29,"doc_title":10,"doc_description":12,"doc_content":134,"file_id":135,"file_url":136,"file_type":137,"file_size":138,"view_count":55,"is_deleted":4,"is_public":22,"is_downloadable":22,"audit_status":22,"page_count":139,"language":140,"language_code":8,"site_id":7,"html_lang":8,"table_of_contents":141,"faqs":142,"seo_title":143,"seo_description":12,"update_tm":80,"read_time":144,"preview_page_count":4},1099513958607,"https://ap-avatar.wpscdn.com/avatar/100002390cf8733938c?x-image-process=image/resize,m_fixed,w_180,h_180&k=1778829742770036399","CulTrace: Tracing Internal Cultural Reasoning in Large Language Models  \nHaeun Yu1 Arnav Arora1 Seogyeong Jeong2 Nadav Borenstein1 Siddhesh Pawar1 Jisu Shin2 Jiho Jin2 Junho Myung2 Alice Oh2 Isabelle Augenstein1  \n1University of Copenhagen 2 KAIST  \n[hayu@di.ku.dk](hayu@di.ku.dk)  \narXiv : 2508 .08879v4 [cs .CL] 2 1 Aug 2026  \nAbstract  \nThe growing deployment of large language models (LLMs) across diverse cultural contexts necessitates a deeper understanding of models’ hidden representations of different cultures.  \nPrior work has evaluated cultural awareness  \nin LLMs by analysing their outputs. This approach overlooks how cultures are represented within the model parameters, missing why models generate incorrect responses. To bridge this gap, we propose CulTrace, a mechanistic interpretability-based method that probes the internal representations of LLMs for cultural knowledge. With CulTrace, we inspect how cultural knowledge is processed across layers and how it is integrated during cultural QA. We find a consistent staged trajectory of cultural reasoning. Models first engage with the question’s domain, then resolve the relevant culture, and finally narrow in on an answer. We also demonstrate that models’ cultural reasoning is imbalanced, showing delayed relevant culture resolution and more confusion with lessrepresented cultures.  \n1 Introduction  \nCultural alignment has become an important aspect of LLM development due to their worldwide deployment. Most existing evaluations of cultural alignment are extrinsic, measuring model outputs on cultural understanding or simulation tasks. While these approaches can reveal when models produce culturally inaccurate responses, they provide limited insight into why, making it challenging to improve performance. To address this gap, we introduce CulTrace, a diagnostic mechanistic interpretability framework that verbalises layer-wise model representations of cultural reasoning into natural language. Using CulTrace, we output chain of reasonings across layers within a model, allowing us to inspect when the model is focusing on various elements pertaining to a query, e.g., the  \nFigure 1: We present an overview of the proposed framework. Given a cultural question about a target culture (Ethiopia), CulTrace decodes the model’s internal representations at each layer to reveal what the model encodes, spanning from topic (e.g., food) and answer candidates (e.g., injera) to cultural context (e.g., misattributed cultures such as Kenya or Japan) .  \ndomain of the question, cultural context, candidate answers and more (Figure 1) .  \nCulTrace (§3) builds on LatentQA (Pan et al., 2026), a method for decoding model activations into natural language by training decoders that take model activations and a question as input and a corresponding answer as output. The trained decoder generalises to tasks beyond training, allowing one to successfully probe the activations for various tasks like recovering hidden system prompts, predicting future generations, or extracting relational knowledge through a decoding question. While Pan et al. (2026) use a specific layer from the model to decode activations, we adapt it by training decoders for each layer and creating a chain of intermediate reasoning across the model’s layers. This helps verbalise how information emerges and is refined across the layers (Table 1) . We apply this framework to cultural QA as it is a challenging task with subjective answers compared to domains like math or code, making it crucial to understand intermediate reasoning and cultural bias.  \nWe employ CulTrace for two cultural probing  \ntasks, tracing intermediate cultural reasoning and identifying the target culture, when answering a cultural question. For the former (§4), we generate intermediate reasoning at each layer and characterise whether the model is processing the question’s domain, cultural context, or answer selection, amongst others. For the latter (§5), we probe for whic","cbCaibyMBtI3wl8x","https://ap.wps.com/l/cbCaibyMBtI3wl8x","pdf",2307998,22,"English","# Introduction\n## Cultural alignment and evaluation gap\n## CulTrace framework overview\n## Probing tasks and datasets","[{\"question\":\"What limitation in existing cultural alignment evaluations does CulTrace address?\",\"answer\":\"Existing evaluations largely analyze model outputs, showing when responses are culturally inaccurate but offering limited insight into why. CulTrace focuses on how cultural information is represented and processed internally.\"},{\"question\":\"How does CulTrace probe cultural reasoning inside an LLM?\",\"answer\":\"CulTrace trains layer-specific decoders to verbalize intermediate model representations. It generates chain-of-reasonings across layers to show when the model engages the domain, resolves cultural context, and selects an answer.\"},{\"question\":\"What key pattern does CulTrace find about cultural reasoning across layers?\",\"answer\":\"It finds a consistent staged trajectory: models first engage the question’s domain, then resolve the relevant culture, and finally narrow in on an answer. Less-represented cultures tend to be resolved later and show more confusion.\"}]","CulTrace - Tracing Internal Cultural Reasoning in Large Language Models - mechanistic interpretability framework | PDF",55]