[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-84982-en":3,"doc-seo-84982-105":29,"detail-sidebar-cat-0-en-105":89},{"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":20,"is_deleted":4,"is_public":21,"is_downloadable":21,"audit_status":21,"page_count":22,"language":23,"language_code":24,"site_id":25,"html_lang":24,"table_of_contents":26,"faqs":27,"seo_title":13,"seo_description":14,"update_tm":28,"read_time":11},84982,7971461740909,"Levi","https://ap-avatar.wpscdn.com/davatar_155a257f0dc6eb9ab79c44ca47cae57d",8,"Research & Report","Interpretable Uncertainty for Adaptive Retrieval and Reasoning in Question Answering","Large language models (LLMs) deliver strong question answering performance but still produce hallucinations and offer limited transparency. Retrieval-augmented generation (RAG) improves factuality, yet deciding when to retrieve and how much to reason is often driven by opaque policies or costly multi-step prompting. The work proposes an uncertainty-aware adaptive QA framework using explicit signals from LLM internal representations, separating knowledge insufficiency from ambiguity or conflict and estimating them efficiently in a single forward pass.","Interpretable Uncertainty for Adaptive Retrieval and Reasoning  \nin Question Answering  \nRitajit Dey  \n[r.dey.1@research.gla.ac.uk](r.dey.1@research.gla.ac.uk)[ ](r.dey.1@research.gla.ac.uk)University of Glasgow Glasgow, UK  \nIadh Ounis  \n[iadh.ounis@glasgow.ac.uk](iadh.ounis@glasgow.ac.uk)[ ](iadh.ounis@glasgow.ac.uk)University of Glasgow Glasgow, UK  \nGraham McDonald  \n[graham.mcdonald@glasgow.ac.uk](graham.mcdonald@glasgow.ac.uk)[ ](graham.mcdonald@glasgow.ac.uk)University of Glasgow Glasgow, UK  \narXiv :2607 .07380v 1 [ cs .IR] 8 Jul 2026  \nAbstract  \nLarge language models (LLMs) achieve a strong performance in question answering (QA), but remain prone to hallucinations and suffer from limited transparency. Retrieval-augmented generation (RAG) can improve factuality, yet decisions about when and how to retrieve from external resources are typically based on opaque policies or computationally inefficient multi-step prompting procedures. We propose an uncertainty-aware framework for adaptive QA based on explicit signals derived from LLM internal representations. We distinguish between knowledge insufficiency and knowledge ambiguity or conflict, and efficiently estimate these from hidden states ina single forward pass. These signals guide system behaviour: RAG is triggered when knowledge is insufficient, while additional reasoning is applied when ambiguity or conflict is high. By grounding adaptive decisions in decomposed and efficiently estimable uncertainty signals, this approach provides a transparent and practical alternative to existing retrieval and reasoning strategies supporting the design of interpretable user-facing tools.  \nCCS Concepts  \n• Information systems → Information retrieval; Question answering; • Computing methodologies → Language models.  \nKeywords  \nQuestion Answering, Retrieval-Augmented Generation, Information Retrieval, Uncertainty Estimation, Large Language Models, Interpretability  \n1 Introduction  \nQuestion answering (QA) is a central task in information retrieval and natural language processing, requiring systems to produce accurate answers to natural language queries. Recent progress has been driven by large language models (LLMs), which generate answers directly by leveraging knowledge acquired during large-scale pretraining [18] . Despite strong empirical performance, such models are prone to hallucinated outputs [8], can rely on incomplete or outdated internal knowledge [13], and suffer from limited transparency in how answers are produced [2, 24] .  \nRetrieval-augmented generation (RAG) [13] has emerged as a dominant paradigm to address these issues by grounding generation in externally retrieved evidence. By decoupling knowledge access from answer generation, RAG-based systems can improve factuality and provide a more interpretable link between outputsand supporting information. However, retrieval is not universally beneficial: for queries where relevant knowledge is already wellrepresented in the LLM, retrieval may introduce noise, increase  \nlatency, and degrade answer quality [14] . This has motivated growing interest in adaptive retrieval strategies that determine whether and how to retrieve external evidence [1, 7, 25] .  \nExisting approaches to adaptive retrieval typically rely on implicit learned decision policies [1, 7, 25], or structured multi-step reasoning-based decision mechanisms [17] . The former are opaque and do not disentangle different sources of uncertainty, while the latter rely on multi-step prompting and are computationally inefficient. As a consequence, it is difficult to expose interpretable, user-facing cues to end-users. This highlights the need for decision mechanisms that are both interpretable and efficient, while explicitly modelling different sources of uncertainty. In QA settings, uncertainty may arise due to knowledge insufficiency, where the model lacks the required information, or due to knowledge ambiguity or conflict, where multiple plausible interpretat","cbCaibNDbnxjuSRh","https://ap.wps.com/l/cbCaibNDbnxjuSRh","pdf",760439,2,1,3,"English","en",105,"# Introduction\n## Problem with hallucinations and limited transparency\n## Adaptive retrieval motivation\n## Distinguishing uncertainty types","[{\"question\":\"How does the proposed approach improve QA compared with standard LLM answering?\",\"answer\":\"It reduces hallucinations and improves transparency by grounding decisions in explicit uncertainty signals derived from LLM internal representations, rather than relying on opaque behaviors.\"},{\"question\":\"What is the difference between knowledge insufficiency and knowledge ambiguity or conflict?\",\"answer\":\"Knowledge insufficiency means the model lacks required information, while ambiguity or conflict occurs when multiple plausible interpretations or competing evidence exist even though relevant facts may be present.\"},{\"question\":\"How does the framework decide when to trigger retrieval versus additional reasoning?\",\"answer\":\"RAG is triggered when knowledge is insufficient, and additional reasoning is applied when uncertainty from ambiguity or conflict is high, using estimates obtained efficiently in a single forward pass.\"}]",1784200011,{"code":4,"msg":30,"data":31},"ok",{"site_id":25,"language":24,"slug":32,"title":13,"keywords":33,"description":14,"schema_data":34,"social_meta":84,"head_meta":86,"extra_data":88,"updated_unix":28},"interpretable-uncertainty-for-adaptive-retrieval-and-reasoning-in-question-answering","",{"@graph":35,"@context":83},[36,51,66],{"@type":37,"itemListElement":38},"BreadcrumbList",[39,43,46,48],{"item":40,"name":41,"@type":42,"position":21},"https://docshare.wps.com","Home","ListItem",{"item":44,"name":45,"@type":42,"position":20},"https://docshare.wps.com/document/","Document",{"item":47,"name":12,"@type":42,"position":22},"https://docshare.wps.com/document/research-report/",{"item":49,"name":13,"@type":42,"position":50},"https://docshare.wps.com/document/interpretable-uncertainty-for-adaptive-retrieval-and-reasoning-in-question-answering/84982/",4,{"url":49,"name":13,"@type":52,"author":53,"headline":13,"publisher":55,"fileFormat":58,"inLanguage":24,"description":14,"dateModified":59,"datePublished":60,"encodingFormat":58,"isAccessibleForFree":61,"interactionStatistic":62},"DigitalDocument",{"name":9,"@type":54},"Person",{"url":40,"name":56,"@type":57},"DocShare","Organization","application/pdf","2026-07-23","2026-07-16",true,{"@type":63,"interactionType":64,"userInteractionCount":20},"InteractionCounter",{"@type":65},"ViewAction",{"@type":67,"mainEntity":68},"FAQPage",[69,75,79],{"name":70,"@type":71,"acceptedAnswer":72},"How does the proposed approach improve QA compared with standard LLM answering?","Question",{"text":73,"@type":74},"It reduces hallucinations and improves transparency by grounding decisions in explicit uncertainty signals derived from LLM internal representations, rather than relying on opaque behaviors.","Answer",{"name":76,"@type":71,"acceptedAnswer":77},"What is the difference between knowledge insufficiency and knowledge ambiguity or conflict?",{"text":78,"@type":74},"Knowledge insufficiency means the model lacks required information, while ambiguity or conflict occurs when multiple plausible interpretations or competing evidence exist even though relevant facts may be present.",{"name":80,"@type":71,"acceptedAnswer":81},"How does the framework decide when to trigger retrieval versus additional reasoning?",{"text":82,"@type":74},"RAG is triggered when knowledge is insufficient, and additional reasoning is applied when uncertainty from ambiguity or conflict is high, using estimates obtained efficiently in a single forward pass.","https://schema.org",{"og:url":49,"og:type":85,"og:title":13,"og:site_name":56,"og:description":14},"article",{"robots":87,"canonical":49},"index,follow",{"doc_id":7,"site_id":25},{"code":4,"msg":5,"data":90},[91,95,99,103,108,113,118,121,126,129,133],{"id":21,"doc_module":4,"doc_module_name":45,"category_name":92,"show_sort_weight":93,"slug":94},"Story & Novel",90,"story-novel",{"id":20,"doc_module":4,"doc_module_name":45,"category_name":96,"show_sort_weight":97,"slug":98},"Literature",80,"literature",{"id":50,"doc_module":4,"doc_module_name":45,"category_name":100,"show_sort_weight":101,"slug":102},"Exam",70,"exam",{"id":104,"doc_module":4,"doc_module_name":45,"category_name":105,"show_sort_weight":106,"slug":107},5,"Comic",60,"comic",{"id":109,"doc_module":4,"doc_module_name":45,"category_name":110,"show_sort_weight":111,"slug":112},6,"Technology",50,"technology",{"id":114,"doc_module":4,"doc_module_name":45,"category_name":115,"show_sort_weight":116,"slug":117},7,"Healthcare",40,"healthcare",{"id":11,"doc_module":4,"doc_module_name":45,"category_name":12,"show_sort_weight":119,"slug":120},30,"research-report",{"id":122,"doc_module":4,"doc_module_name":45,"category_name":123,"show_sort_weight":124,"slug":125},9,"Religion & Spirituality",20,"religion-spirituality",{"id":124,"doc_module":4,"doc_module_name":45,"category_name":127,"show_sort_weight":124,"slug":128},"World Cup","world-cup",{"id":130,"doc_module":4,"doc_module_name":45,"category_name":131,"show_sort_weight":130,"slug":132},10,"Lifestyle","lifestyle",{"id":134,"doc_module":4,"doc_module_name":45,"category_name":135,"show_sort_weight":104,"slug":136},19,"General","general"]