[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-84527-en":3,"doc-seo-84527-105":29,"detail-sidebar-cat-0-en-105":91},{"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":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},84527,962075006959,"Anda","https://ap-avatar.wpscdn.com/avatar/e0002397efbe92a78e?_k=1776741047341049297",8,"Research & Report","Readable but Not Controllable Neuron-Level Evidence for Medical LLM Hallucination","Medical LLM hallucination remains a major barrier to safe deployment, and even detectable hallucination signals may not be controllable. This paper studies four open-source medical QA models across multiple datasets using conditioned neuron probes with AUROC 0.77–0.86. The hallucination-related representation is distributed and redundant: random neuron subsets of a few hundred and low-dimensional projections preserve most detection power. Causal steering reveals a sharp detection–controllability gap across 16 model–dataset settings, indicating legible but not reliably actionable representations.","arXiv :2607 .00158v1 [ cs .CL] 30 Jun 2026  \nReadable but Not Controllable: Neuron-Level Evidence for Medical LLM Hallucination  \nVijay Vankadaru1, Asha Matthews1, Tanya Roosta∗,1 & Peyman Passban ∗  \n1 School of Information University of California, Berkeley California, USA  \n[vankadaruvijay@ischool.berkeley.edu](vankadaruvijay@ischool.berkeley.edu) , [passban.peyman@gmail.com](passban.peyman@gmail.com)  \nAbstract  \nHallucination remains one of the central obstacles to deploying medical LLMs. Yet, even when hallucination can be detected, it is still unclear whether the internal representations associated with it can be used for control rather than detection alone. Using four open-source models across a suite of medical question-answering datasets, we show that a simple, carefully conditioned probe can reliably detect hallucination, with AUROC scores between 0.77 and 0.86 in our case. We further show that this signal is distributed and redundant rather than narrowly localized. Systematically selected neurons outperform random neurons only at very small subset sizes, whereas random subsets of a few hundred neurons recover nearly the full signal, and low-dimensional random projections preserve most of the detection performance. Beyond detection, we test whether this representation is causally actionable. Across 16 model–dataset combinations, our results reveal a sharp gap between decodability and controllability.  \nThe same internal structure that makes hallucination easy to detect does not translate into reliable neuron-level control. These findings show that medical hallucination seems to be readily visible in internal activations, but not easily corrected by steering the neurons most associated with it. More broadly, our results suggest that hallucination mitigation is not simply a matter of identifying the right neurons, and point to a deeper separation between what representations reveal and what they allow us to change.  \n1 Introduction  \nThis paper studies whether factual hallucination in medical LLMs can be understood through the lens of hallucination-associated neurons. Specifically, we investigate a mechanism for detecting neurons that may help differentiate truthful answers from hallucinated ones. Once these neurons are selected, we can use them to classify internal patterns and behaviors of LLMs and, hopefully, intervene in their hallucinations. For this purpose, we simply measure the contribution of each neuron to the output of its associated layer.  \nA common technique to measure this type of contribution is CETT (Zhang et al., 2024), which is (almost) equivalent to the magnitude of a neuron and shows how much a neuron contributes to the model’s generated tokens. The closest paper that relies on a CETT-style analysis is H-Neuron (Gao et al., 2025) . In that work, the authors identify neurons whose contribution patterns are strongly linked to hallucinated answers. In this respect, our work follows a similar direction. However, we ask a more specific question. Are hallucinationassociated activations in medical models merely detectable, or can they also serve as causal control levers? In doing so, we move beyond H-Neuron and investigate the possibility of causal control.  \n∗ This work was conducted independently and is not associated with the author’s employer. All views and conclusions are solely those of the author.  \nH-Neuron suggests that some behaviors, such as over-compliance, are linked to a small set of important neurons and can show causal effects when those neurons are changed. In contrast, this project studies factual hallucination in medical LLMs and finds that the signal is spread across hundreds of neurons. The signal can be detected, but changing those neurons does not necessarily change the model’s hallucination behavior. Our experimental results attempt to dissect why this happens. Throughout the paper, further differences among the problems we investigate and others will become clearer as we provide ","cbCaitN2hCCtHlKD","https://ap.wps.com/l/cbCaitN2hCCtHlKD","pdf",561295,1,15,"English","en",105,"# Introduction\n# Background","[{\"question\":\"How does the paper detect hallucinations in medical LLMs?\",\"answer\":\"It uses a carefully conditioned neuron probe that measures how each neuron contributes to the output of its associated layer, achieving AUROC scores between 0.77 and 0.86 across datasets and models.\"},{\"question\":\"What does the paper find about the representation of hallucination signals in neurons?\",\"answer\":\"Hallucination signals are distributed and redundant rather than narrowly localized; random subsets of a few hundred neurons can recover nearly the full detection signal, and low-dimensional random projections retain most performance.\"},{\"question\":\"Why are neuron activations detectable but not controllable?\",\"answer\":\"Across 16 model–dataset combinations, the same internal structure that supports strong detectability via probing fails to provide reliable causal control when neurons are steered, revealing a sharp gap between decodability and controllability.\"}]",1784196410,38,{"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":86,"head_meta":88,"extra_data":90,"updated_unix":27},"readable-but-not-controllable-neuron-level-evidence-for-medical-llm-hallucination","",{"@graph":35,"@context":85},[36,53,68],{"@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/readable-but-not-controllable-neuron-level-evidence-for-medical-llm-hallucination/84527/",4,{"url":51,"name":13,"@type":54,"author":55,"headline":13,"publisher":57,"fileFormat":60,"inLanguage":23,"description":14,"dateModified":61,"datePublished":62,"encodingFormat":60,"isAccessibleForFree":63,"interactionStatistic":64},"DigitalDocument",{"name":9,"@type":56},"Person",{"url":40,"name":58,"@type":59},"DocShare","Organization","application/pdf","2026-07-17","2026-07-16",true,{"@type":65,"interactionType":66,"userInteractionCount":20},"InteractionCounter",{"@type":67},"ViewAction",{"@type":69,"mainEntity":70},"FAQPage",[71,77,81],{"name":72,"@type":73,"acceptedAnswer":74},"How does the paper detect hallucinations in medical LLMs?","Question",{"text":75,"@type":76},"It uses a carefully conditioned neuron probe that measures how each neuron contributes to the output of its associated layer, achieving AUROC scores between 0.77 and 0.86 across datasets and models.","Answer",{"name":78,"@type":73,"acceptedAnswer":79},"What does the paper find about the representation of hallucination signals in neurons?",{"text":80,"@type":76},"Hallucination signals are distributed and redundant rather than narrowly localized; random subsets of a few hundred neurons can recover nearly the full detection signal, and low-dimensional random projections retain most performance.",{"name":82,"@type":73,"acceptedAnswer":83},"Why are neuron activations detectable but not controllable?",{"text":84,"@type":76},"Across 16 model–dataset combinations, the same internal structure that supports strong detectability via probing fails to provide reliable causal control when neurons are steered, revealing a sharp gap between decodability and controllability.","https://schema.org",{"og:url":51,"og:type":87,"og:title":13,"og:site_name":58,"og:description":14},"article",{"robots":89,"canonical":51},"index,follow",{"doc_id":7,"site_id":24},{"code":4,"msg":5,"data":92},[93,97,101,105,110,115,120,123,128,131,135],{"id":20,"doc_module":4,"doc_module_name":45,"category_name":94,"show_sort_weight":95,"slug":96},"Story & Novel",90,"story-novel",{"id":46,"doc_module":4,"doc_module_name":45,"category_name":98,"show_sort_weight":99,"slug":100},"Literature",80,"literature",{"id":52,"doc_module":4,"doc_module_name":45,"category_name":102,"show_sort_weight":103,"slug":104},"Exam",70,"exam",{"id":106,"doc_module":4,"doc_module_name":45,"category_name":107,"show_sort_weight":108,"slug":109},5,"Comic",60,"comic",{"id":111,"doc_module":4,"doc_module_name":45,"category_name":112,"show_sort_weight":113,"slug":114},6,"Technology",50,"technology",{"id":116,"doc_module":4,"doc_module_name":45,"category_name":117,"show_sort_weight":118,"slug":119},7,"Healthcare",40,"healthcare",{"id":11,"doc_module":4,"doc_module_name":45,"category_name":12,"show_sort_weight":121,"slug":122},30,"research-report",{"id":124,"doc_module":4,"doc_module_name":45,"category_name":125,"show_sort_weight":126,"slug":127},9,"Religion & Spirituality",20,"religion-spirituality",{"id":126,"doc_module":4,"doc_module_name":45,"category_name":129,"show_sort_weight":126,"slug":130},"World Cup","world-cup",{"id":132,"doc_module":4,"doc_module_name":45,"category_name":133,"show_sort_weight":132,"slug":134},10,"Lifestyle","lifestyle",{"id":136,"doc_module":4,"doc_module_name":45,"category_name":137,"show_sort_weight":106,"slug":138},19,"General","general"]