[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"detail-sidebar-cat-0-en-105":3,"doc-seo-201215-105":59,"doc-detail-201215-en":131},{"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":124,"head_meta":126,"extra_data":128,"updated_unix":130},105,"en","evaluating-the-casual-reasoning-abilities-of-large-language-models-causal-reasoning-dataset-and-explanation-test-suite","Evaluating the Casual Reasoning Abilities of Large Language Models - Causal REasoning Dataset and Explanation Test Suite","","Explores how large language models produce seemingly step-by-step explanations and assesses whether such “reasoning” reflects genuine causal thinking. Argues that explainability requires eliciting model reasoning, then evaluating the faithfulness and accuracy of that stated thinking. Introduces CReDETS, a causal reasoning dataset built from LSAT logic games with 442 hand-annotated explanations. Benchmarks GPT-3, GPT-3.5, and GPT-4 over 10 trials, finding low accuracy and limited human-level performance despite improved consistency.",{"@graph":69,"@context":123},[70,84,106],{"@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/evaluating-the-casual-reasoning-abilities-of-large-language-models-causal-reasoning-dataset-and-explanation-test-suite/201215/",{"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/evaluating-the-casual-reasoning-abilities-of-large-language-models-causal-reasoning-dataset-and-explanation-test-suite/201215.png","ImageObject",300,407,{"name":92,"@type":93},"Maya Linwood","Person",{"url":74,"name":95,"@type":96},"DocShare","Organization","application/pdf","2026-11-05","2026-09-04",true,{"@type":102,"interactionType":103,"userInteractionCount":105},"InteractionCounter",{"@type":104},"ViewAction",24,{"@type":107,"mainEntity":108},"FAQPage",[109,115,119],{"name":110,"@type":111,"acceptedAnswer":112},"Why is it urgent to study LLM reasoning processes?","Question",{"text":113,"@type":114},"LLMs can behave dangerously when users treat them as “oracles,” and their explanations may be unreliable mimicry. The work argues that society-impacting integration requires insight into the model’s step-by-step reasoning.","Answer",{"name":116,"@type":111,"acceptedAnswer":117},"What is CReDETS and what makes it different?",{"text":118,"@type":114},"CReDETS is the Causal REasoning Dataset and Explanation Test Suite, built from LSAT logic game questions. It includes hand-written explanations for each question, not just answers and choices.",{"name":120,"@type":111,"acceptedAnswer":121},"How were GPT-3, GPT-3.5, and GPT-4 evaluated on CReDETS?",{"text":122,"@type":114},"Each of the 442 questions was run via separate API calls, and the evaluation was repeated for 10 trials. Results report trial averages and detailed trial accuracies.","https://schema.org",{"og:url":83,"og:type":125,"og:title":65,"og:site_name":95,"og:description":67},"article",{"robots":127,"canonical":83},"index,follow",{"doc_id":129,"site_id":62},201215,1788527604,{"code":4,"msg":5,"data":132},{"doc_id":129,"user_id":133,"nickname":92,"user_avatar":134,"doc_module":4,"category_id":39,"category_name":40,"doc_title":65,"doc_description":67,"doc_content":135,"file_id":136,"file_url":137,"file_type":138,"file_size":139,"view_count":105,"is_deleted":4,"is_public":8,"is_downloadable":8,"audit_status":8,"page_count":8,"language":140,"language_code":63,"site_id":62,"html_lang":63,"table_of_contents":141,"faqs":142,"seo_title":143,"seo_description":67,"update_tm":130,"read_time":81,"preview_page_count":8,"preview_urls":144},962084928432,"https://ap-avatar.wpscdn.com/davatar_155a257f0dc6eb9ab79c44ca47cae57d","Evaluating the Casual Reasoning Abilities of Large Language Models  \nIsha Puri, Hima Lakkaraju  \nHarvard University  \nIntroduction CoFrNets Analysis  \nCurrent States: Despite their impressive abilities, however, Large Language Models suffer from several behaviors that can prove dangerous to users treating them as ”oracles” and suggest that any ”reasoning” demonstrated by the models is not genuine and simply an unreliable mimicry of large-scale training data. As these models are being integrated into impactful parts of society, it is thus urgent that we gain insight into the step-by-step \"reasoning processes\" taken by LLMs to arrive at their outputs.  \nA new paradigm for \"Explainability\"  \nWe argue that explainability in language models is about extracting reasoning from the model, whether through prompting [?], one shot [?]/few shot learning [?], etc. Reasoning is a cognitive process that involves drawing conclusions based on available information, often through logical steps or inferences. By focusing on eliciting the model’s reasoning, we can better understand how a model processes and manipulates information to arrive at its conclusions. This approach allows us to extract insights into the model’s internal decision-making processes, thus enabling a deeper understanding of the model’s behavior.  \nIn order to fully trust that the reasoning a LLM outputs for a given prompt can be used as evidence for its true internal workings, however, we need to have faith in the reasoning abilities of language models. We need to measure how well LLMs can explain their \"thinking\" -how accurate is their reasoning?  \nIntroduction to CReDETS  \nAlthough there are datasets such as LogiQA for measuring general logic abilities of large language models through open ended question and answering benchmarks, there is lack of data sources that explicitly focus on complex causal reasoning Q&A and include high quality explanations of those answers as well.  \nTo this end, we introduce CReDETS, the Causal REasoning Dataset and Explanation Test Suite, a novel, first-of-its-kind causal reasoning dataset with hand-annotated explanations.  \nWe hope that the introduction of this dataset will allow researchers to continue to evaluate and improve the reasoning abilities of various generations of language models.  \nThese questions are based on the LSAT, which is one of the only professional tests that doesn’t require any subject matter knowledge, and thus is a perfect basis for a causal reasoning dataset. These professional exam questions are written by philosophy and logic experts to specifically measure causal reasoning ability.  \nWe curated 442 samples, each of which is based on a premise involving a set of characters and rules that define relationships between them.  \nFigure 1:Structure of LSAT Logic Games Questions  \nStructure of CReDETS  \nFor each question, we include not only the question, answer choices, and correct answer, but also a hand-written explanation for each question, a unique differentiation of our dataset with respect to all others in the field such a LogiQA. This tests the capabilities of language models to not just choose the right answer options (MCQ) but also to explain reasoning for each question.  \nFigure 2:CReDETS Dataset-Distribution of Question Categories (Total Questions: 442)  \nBenchmarking Results  \nIn order to measure the accuracy of our three test models-GPT3, GPT3.5, and GPT4-on the questions in the CReDETS dataset, we ran 10 trials of the 442 questions. Each question was run via a separate API call. The results can be seen in table 1 below.  \n\n|  | Model |  |  |\n| --- | --- | --- | --- |\n|  | GPT 3 | GPT 3.5 | GPT 4 |\n| Trial Average | 0.198 | 0.207 | 0.248 |\n| Trial 1 | .199 | .205 | .278 |\n| Trial 2 | .201 | .210 | .282 |\n| Trial 3 | .192 | .212 | .291 |\n| Trial 4 | .196 | .203 | .271 |\n| Trial 5 | .208 | .213 | .269 |\n| Trial 6 | .205 | .210 | .264 |\n| Trial 7 | .199 | .201 | .280 |\n| Trial 8 | .199 | .199 | .271 |\n| Trial 9 |","cbCais933ag1lsXM","https://ap.wps.com/l/cbCais933ag1lsXM","pdf",382951,"English","# Introduction\n## Current States\n## A new paradigm for \"Explainability\"\n## Introduction to CReDETS\n## Structure of CReDETS\n# Benchmarking Results","[{\"question\":\"Why is it urgent to study LLM reasoning processes?\",\"answer\":\"LLMs can behave dangerously when users treat them as “oracles,” and their explanations may be unreliable mimicry. The work argues that society-impacting integration requires insight into the model’s step-by-step reasoning.\"},{\"question\":\"What is CReDETS and what makes it different?\",\"answer\":\"CReDETS is the Causal REasoning Dataset and Explanation Test Suite, built from LSAT logic game questions. It includes hand-written explanations for each question, not just answers and choices.\"},{\"question\":\"How were GPT-3, GPT-3.5, and GPT-4 evaluated on CReDETS?\",\"answer\":\"Each of the 442 questions was run via separate API calls, and the evaluation was repeated for 10 trials. Results report trial averages and detailed trial accuracies.\"}]","Evaluating the Casual Reasoning Abilities of Large Language Models - Causal REasoning Dataset and Explanation Test Suite | PDF",[145],{"page_no":8,"url":146,"width":147,"height":148},"https://res-academy.cache.wpscdn.com/upload/kdocshare/2026-11-08/80363a60fbe1fab7fb4f6a56f1e50dcc.webp",983,552]