[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-84636-en":3,"doc-seo-84636-105":29,"detail-sidebar-cat-0-en-105":90},{"code":4,"msg":5,"data":6},0,"success",{"doc_id":7,"user_id":8,"nickname":9,"user_avatar":10,"doc_module":4,"category_id":11,"category_name":12,"doc_title":13,"doc_description":14,"doc_content":15,"file_id":16,"file_url":17,"file_type":18,"file_size":19,"view_count":4,"is_deleted":4,"is_public":20,"is_downloadable":20,"audit_status":20,"page_count":21,"language":22,"language_code":23,"site_id":24,"html_lang":23,"table_of_contents":25,"faqs":26,"seo_title":13,"seo_description":14,"update_tm":27,"read_time":28},84636,3848291630094,"Emma Wilson","https://eur-avatar.wpscdn.com/davatar_085a072bc5b1113ac321206ff7593b45",8,"Research & Report","CAUSALSTEWARD: An Agentic Divide-Conquer-Combine Copilot for Causal Discovery","Learning causal models from high-dimensional data remains challenging when real-world violations of identifiability assumptions prevent reliable causal discovery. Although large volumes of prior knowledge exist in documents and web text, integrating it effectively into causal discovery is still unresolved. CausalSTeward (CAST) introduces a human-in-the-loop multi-agent framework that interactively assembles large causal models via divide-and-conquer partitioning. CAST fuses retrieved prior knowledge using retrieval augmented generation, conditional independence tests, and expert feedback to improve causal reasoning quality, trustworthiness, and robustness.","arXiv :2607 .0 1936v 1 [ cs .MA] 2 Jul 2026  \nCAUSALSTEWARD: AN AGENTIC DIVIDE-CONQUERCOMBINE COPILOT FOR CAUSAL DISCOVERY  \nNicholas Tagliapietra 1 ,2 ,†,B, Gian Lorenzo Marchioni5 ,†,⋆, Moritz Willig 1 , Juergen Luettin2 , Lavdim Halilaj2 , Kristian Kersting 1 ,3 ,4  \n1 : Computer Science Department, TU Darmstadt, Germany  \n2 : Bosch Center for Artificial Intelligence, Renningen, Germany  \n3 : Hessian Center for AI (hessian.AI), Germany  \n4 : German Research Center for AI (DFKI), Germany  \n5 : LUISS University, Rome, Italy  \nABSTRACT  \nLearning causal models from high-dimensional data is a significant challenge, particularly in real-world settings where violations of core assumptions lead to causal identifiability issues. Although massive amounts of prior knowledge are available, and contain valuable causal information, effectively integrating this knowledge into the causal discovery process remains an open problem. We introduce CausalSTeward (CAST), a novel humanin-the-loop framework for interactively assembling large causal models. CausalSteward is a multi-agent collaborative system that tackles high-dimensional causality through a divide-and-conquer approach where large clusters of variables are iteratively partitioned and then separately analyzed. Our framework fuses prior knowledge with a data-driven approach by using tailored tools such as retrieval augmented generation and conditional independence tests. Finally, we use this work to examine the capabilities and limitations of causal reasoning in multi-agent frameworks, and how the human-in-the-loop can contribute to accurate and trustworthy results.  \n1 INTRODUCTION  \nModern science and engineering increasingly demand for causal models (Pearl, 2009) that move beyond purely statistical correlation to predict the outcomes of manipulations on the system (interventions) and reasoning about hypothetical scenarios (counterfactuals) . Performing such inferences often requires a causal graph representing the network of cause-and-effect relationships, however the true causal graph is rarely known. Existing approaches (Spirtes et al., 2001; Shimizu et al., 2011; Spirtes, 2001) commonly attempt to learn this graph from data by relying on causal discovery methods. However, causal discovery from observational data is fundamentally limited by the assumptions required for causal identifiability (Shimizu et al., 2011; Zhang & Hyvarinen, 2012; Jaber et al., 2018) . Violations of these assumptions can result in multiple causal graphs being statistically undistinguishable, which leads to non-identifiable causal relationships and limits accurate causal discovery. These limitations can be overcome by integrating prior knowledge (O’Donnell et al., 2006) . Domain experts can manually incorporate their knowledge by imposing edge constraints, but this has limited utility in high-dimensional settings (Constantinou et al., 2023) . In particular,  \n†Equal Contribution.  \nB Correspondence to: [tagliapietra.nicholas@gmail.com](tagliapietra.nicholas@gmail.com)  \n⋆Work carried on during an internship at the Bosch Center for Artificial Intelligence.  \nCaST  \nObservational Data  \nDomain Knowledge  \n| \u003Cbr>0. Explain\u003Cbr>\u003Cbr>\u003Cbr>User\u003Cbr>It means that the patient had an Xray.\u003Cbr>\u003Cbr>Explainer\u003Cbr>Done. Here is a detailed description of the dataset: …. | \u003Cbr>1. Divide\u003Cbr> | \u003Cbr>2. Conquer\u003Cbr>\u003Cbr>\u003Cbr>User\u003Cbr>Yes, I confirm! | \u003Cbr>3. Combine\u003Cbr>\u003Cbr>\u003Cbr>User\u003Cbr>Yes, those connections are valid! |\n| --- | --- | --- | --- |\n\nFigure 1: CausalSteward (CAST): This paper develops and showcases the utility of a human-in-the-loop, multi-agent system for causal discovery, leveraging the combined strength of automated discovery from data and prior knowledge of LLM, together with human-expert feedback to incorporate domain specific knowledge.  \nunstructured text is a rich source of prior knowledge that is dense of causal information, while being massively available in the web, documents, and more. Recent works using Large","cbCaimLHW0LcPmhk","https://ap.wps.com/l/cbCaimLHW0LcPmhk","pdf",1390963,1,61,"English","en",105,"# Introduction\n## Motivation and Identifiability Limits\n## Integrating Prior Knowledge and LLM Guidance\n# CausalSTeward (CAST) Framework\n## Divide-and-Conquer Partitioning\n## Local Graph Estimation and Merging\n# Experiments and Evaluation\n## Datasets and Baselines\n## Ablation Findings","[{\"question\":\"Why is causal discovery from observational data difficult in high-dimensional real-world settings?\",\"answer\":\"Causal discovery relies on assumptions required for causal identifiability. When these assumptions are violated, multiple causal graphs can become statistically indistinguishable, making causal relationships non-identifiable.\"},{\"question\":\"How does CausalSTeward (CAST) integrate prior knowledge with data-driven discovery?\",\"answer\":\"CAST retrieves prior knowledge through human-in-the-loop interaction and retrieval augmented generation, then uses that knowledge alongside conditional independence tests on observational data to estimate local causal graphs.\"},{\"question\":\"What is the role of the human-in-the-loop in CAST’s performance?\",\"answer\":\"CAST is human-first: it assists experts in building a causal model while automating prior knowledge collection. Ablations disabling retrieval augmented generation and the human-in-the-loop show strong performance depends on using both components jointly.\"}]",1784197378,154,{"code":4,"msg":30,"data":31},"ok",{"site_id":24,"language":23,"slug":32,"title":13,"keywords":33,"description":14,"schema_data":34,"social_meta":85,"head_meta":87,"extra_data":89,"updated_unix":27},"causalsteward-an-agentic-divide-conquer-combine-copilot-for-causal-discovery","",{"@graph":35,"@context":84},[36,53,67],{"@type":37,"itemListElement":38},"BreadcrumbList",[39,43,47,50],{"item":40,"name":41,"@type":42,"position":20},"https://docshare.wps.com","Home","ListItem",{"item":44,"name":45,"@type":42,"position":46},"https://docshare.wps.com/document/","Document",2,{"item":48,"name":12,"@type":42,"position":49},"https://docshare.wps.com/document/research-report/",3,{"item":51,"name":13,"@type":42,"position":52},"https://docshare.wps.com/document/causalsteward-an-agentic-divide-conquer-combine-copilot-for-causal-discovery/84636/",4,{"url":51,"name":13,"@type":54,"author":55,"headline":13,"publisher":57,"fileFormat":60,"inLanguage":23,"description":14,"dateModified":61,"datePublished":61,"encodingFormat":60,"isAccessibleForFree":62,"interactionStatistic":63},"DigitalDocument",{"name":9,"@type":56},"Person",{"url":40,"name":58,"@type":59},"DocShare","Organization","application/pdf","2026-07-16",true,{"@type":64,"interactionType":65,"userInteractionCount":4},"InteractionCounter",{"@type":66},"ViewAction",{"@type":68,"mainEntity":69},"FAQPage",[70,76,80],{"name":71,"@type":72,"acceptedAnswer":73},"Why is causal discovery from observational data difficult in high-dimensional real-world settings?","Question",{"text":74,"@type":75},"Causal discovery relies on assumptions required for causal identifiability. When these assumptions are violated, multiple causal graphs can become statistically indistinguishable, making causal relationships non-identifiable.","Answer",{"name":77,"@type":72,"acceptedAnswer":78},"How does CausalSTeward (CAST) integrate prior knowledge with data-driven discovery?",{"text":79,"@type":75},"CAST retrieves prior knowledge through human-in-the-loop interaction and retrieval augmented generation, then uses that knowledge alongside conditional independence tests on observational data to estimate local causal graphs.",{"name":81,"@type":72,"acceptedAnswer":82},"What is the role of the human-in-the-loop in CAST’s performance?",{"text":83,"@type":75},"CAST is human-first: it assists experts in building a causal model while automating prior knowledge collection. Ablations disabling retrieval augmented generation and the human-in-the-loop show strong performance depends on using both components jointly.","https://schema.org",{"og:url":51,"og:type":86,"og:title":13,"og:site_name":58,"og:description":14},"article",{"robots":88,"canonical":51},"index,follow",{"doc_id":7,"site_id":24},{"code":4,"msg":5,"data":91},[92,96,100,104,109,114,119,122,127,130,134],{"id":20,"doc_module":4,"doc_module_name":45,"category_name":93,"show_sort_weight":94,"slug":95},"Story & Novel",90,"story-novel",{"id":46,"doc_module":4,"doc_module_name":45,"category_name":97,"show_sort_weight":98,"slug":99},"Literature",80,"literature",{"id":52,"doc_module":4,"doc_module_name":45,"category_name":101,"show_sort_weight":102,"slug":103},"Exam",70,"exam",{"id":105,"doc_module":4,"doc_module_name":45,"category_name":106,"show_sort_weight":107,"slug":108},5,"Comic",60,"comic",{"id":110,"doc_module":4,"doc_module_name":45,"category_name":111,"show_sort_weight":112,"slug":113},6,"Technology",50,"technology",{"id":115,"doc_module":4,"doc_module_name":45,"category_name":116,"show_sort_weight":117,"slug":118},7,"Healthcare",40,"healthcare",{"id":11,"doc_module":4,"doc_module_name":45,"category_name":12,"show_sort_weight":120,"slug":121},30,"research-report",{"id":123,"doc_module":4,"doc_module_name":45,"category_name":124,"show_sort_weight":125,"slug":126},9,"Religion & Spirituality",20,"religion-spirituality",{"id":125,"doc_module":4,"doc_module_name":45,"category_name":128,"show_sort_weight":125,"slug":129},"World Cup","world-cup",{"id":131,"doc_module":4,"doc_module_name":45,"category_name":132,"show_sort_weight":131,"slug":133},10,"Lifestyle","lifestyle",{"id":135,"doc_module":4,"doc_module_name":45,"category_name":136,"show_sort_weight":105,"slug":137},19,"General","general"]