[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"detail-sidebar-cat-0-en-105":3,"doc-seo-273708-105":59,"doc-detail-273708-en":130},{"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":123,"head_meta":125,"extra_data":127,"updated_unix":129},105,"en","inhibitory-attention-for-clinical-longcontext-reasoning-characterizing-and-mitigating-lost-in-the-middle-effects-in-ehr-processing","INHIBITORY ATTENTION FOR CLINICAL LONGCONTEXT REASONING - Characterizing and Mitigating Lost-in-the-Middle Effects in EHR Processing","","Electronic health records routinely exceed 100,000 tokens per patient, yet large language models show systematic positional retrieval bias: the lost-in-the-middle (LitM) effect. For clinical notes, omitted or misplaced facts can directly affect care, including outdated medications, mid-record lab values, and contradicting diagnoses. The work defines the clinical lost-in-the-middle (CLitM) problem, empirically characterizes it across clinical fact domains, and evaluates QCCS as a query-conditioned suppression remedy versus multiple retrieval and reranking baselines.",{"@graph":69,"@context":122},[70,84,105],{"@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/inhibitory-attention-for-clinical-longcontext-reasoning-characterizing-and-mitigating-lost-in-the-middle-effects-in-ehr-processing/273708/",{"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/inhibitory-attention-for-clinical-longcontext-reasoning-characterizing-and-mitigating-lost-in-the-middle-effects-in-ehr-processing/273708.png","ImageObject",300,407,{"name":92,"@type":93},"Skyler","Person",{"url":74,"name":95,"@type":96},"DocShare","Organization","application/pdf","2026-10-31","2026-09-15",true,{"@type":102,"interactionType":103,"userInteractionCount":52},"InteractionCounter",{"@type":104},"ViewAction",{"@type":106,"mainEntity":107},"FAQPage",[108,114,118],{"name":109,"@type":110,"acceptedAnswer":111},"What is the lost-in-the-middle (LitM) effect in long-context language model retrieval?","Question",{"text":112,"@type":113},"Information located near the center of a long input context is retrieved far less reliably than information near the edges, producing a systematic positional accuracy bias.","Answer",{"name":115,"@type":110,"acceptedAnswer":116},"Why is the clinical version of LitM (CLitM) especially important for EHR use?",{"text":117,"@type":113},"Clinical facts in the middle of a patient record, such as outdated medications, critical lab values, or contradicting diagnoses, may represent the most consequential information for decision support.",{"name":119,"@type":110,"acceptedAnswer":120},"How does Query-Conditioned Clinical Suppression (QCCS) address the problem?",{"text":121,"@type":113},"QCCS uses a lightweight character n-gram, query-conditioned context selection gate to suppress unhelpful content and improve instruction-following accuracy compared with retrieval-only and reranking comparators.","https://schema.org",{"og:url":83,"og:type":124,"og:title":65,"og:site_name":95,"og:description":67},"article",{"robots":126,"canonical":83},"index,follow",{"doc_id":128,"site_id":62},273708,1789472597,{"code":4,"msg":5,"data":131},{"doc_id":128,"user_id":132,"nickname":92,"user_avatar":133,"doc_module":4,"category_id":39,"category_name":40,"doc_title":65,"doc_description":67,"doc_content":134,"file_id":135,"file_url":136,"file_type":137,"file_size":138,"view_count":52,"is_deleted":4,"is_public":8,"is_downloadable":8,"audit_status":8,"page_count":139,"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":129,"read_time":144,"preview_page_count":4},2336464648746,"https://ap-avatar.wpscdn.com/davatar_276721f389ce27ea32af1340a28f341c","arXiv : 2608 .20348v 1 [cs .CL] 15 Jun 2026  \nINHIBITORY ATTENTION FOR CLINICAL LONGCONTEXT REASONING:  \nCHARACTERIZING AND MITIGATING LOST-IN-THEMIDDLE EFFECTS IN EHR PROCESSING  \nSanjay Basu  \nDepartment of Medicine, University of California San Francisco, San Francisco, CAWaymark, San Francisco, CA  \n[sanjay.basu@ucsf.edu](sanjay.basu@ucsf.edu)  \nABSTRACT  \nElectronic health records now routinely exceed 100,000 tokens per patient. Yet large language models exhibit a systematic positional retrieval bias, the lost-inthe-middle (LitM) effect, in which information located near the center of a long input context is retrieved far less reliably than information near the edges. For clinical applications, this bias is not a minor benchmark quirk: a discontinued medication recorded two years ago, a critical lab value from the middle of a patient’s record, or an old diagnosis that contradicts a current treatment plan may represent the most clinically consequential fact in the entire note. We term this the clinical lost-in-the-middle (CLitM) problem, provide its first systematic empirical characterization across clinical fact domains using MedAlign, and compare context selection strategies as remedies. Across 2,196 instruction-response pairs and six language models, we observe a 21.9 percentage-point gap between peak accuracy (59.5%[95% clustered CI: 46.3, 71.0] at the 20–30% position decile) and trough accuracy (37.6%[23.2, 52.5] at 70–80%; gap clustered CI: [17.7, 45.2] pp);  \n67.8% of reference answers reside between the 10th and 90th percentiles of the EHR timeline, placing them within the CLitM accuracy trough. We introduce Query-Conditioned Clinical Suppression (QCCS), a lightweight character n-gram query-conditioned context selection gate, and evaluate it against BM25, BM25 with section-header filtering, dense retrieval (sentence-transformers), and crossencoder reranking in a multi-arm proof-of-concept experiment (N=83 held-out instructions) . End-to-end instruction-following experiments with Qwen2.5-7BInstruct (16k-token context window) show that QCCS substantially outperforms all five comparators, evaluated by LLM-as-judge semantic scoring (primary: Claude Haiku; second judge: claude-sonnet-4-6; Cohen’s κ=0 .767 for QCCS, the arm with meaningful positive-class variation): QCCS 16.7%[95% CI: 3, 30] vs. BM25 3.3%[0, 10] vs. cross-encoder 0 .0%[0, 0] vs. dense 0 .0%[0, 0] vs. full context 6.7%[0, 17] for middle-position instructions; overall, QCCS reaches 25.3%[17, 35] vs. ≤3.6% for all retrieval-only comparators. This end-to-end advantage is not attributable to higher retrieval recall: at k=20, BM25 retrieves the gold-standard evidence sentence in 98.8% of test instructions (QCCS: 34.9%); cross-encoder reranking achieves 96.4% overall; dense retrieval 94.0% . A conditional accuracy analysis reveals that Stage-2 accuracy remains ≤2.6% for all retrieval-based comparators even in the 74–78 instructions where those arms successfully retrieve the gold-evidence sentence (by lexical overlap), while QCCS achieves 25.0% accuracy in the 52 instructions where it does not retrieve the gold sentence, establishing that, in this proof-of-concept evaluation, query-aligned context selection is a stronger predictor of EHR instruction-following accuracy than gold-sentence retrieval recall.  \nA no-query ablation (Appendix H) confirms that query conditioning accounts for +23.8 percentage points of middle-position recall. Code is available at [https:](https:)//[github.com/sanjaybasu/inhibitory-attention-ehr](github.com/sanjaybasu/inhibitory-attention-ehr).  \n1 INTRODUCTION  \nModern electronic health records are among the longest structured documents routinely encountered by language models. A single longitudinal patient record from an integrated delivery network can exceed 100,000 tokens when all encounter notes, laboratory results, medication histories, and social history entries are concatenated in chronological order. Recent advances in transforme","cbCais2DjKaEO0i5","https://ap.wps.com/l/cbCais2DjKaEO0i5","pdf",1248629,29,"English","# Abstract\n# Introduction","[{\"question\":\"What is the lost-in-the-middle (LitM) effect in long-context language model retrieval?\",\"answer\":\"Information located near the center of a long input context is retrieved far less reliably than information near the edges, producing a systematic positional accuracy bias.\"},{\"question\":\"Why is the clinical version of LitM (CLitM) especially important for EHR use?\",\"answer\":\"Clinical facts in the middle of a patient record, such as outdated medications, critical lab values, or contradicting diagnoses, may represent the most consequential information for decision support.\"},{\"question\":\"How does Query-Conditioned Clinical Suppression (QCCS) address the problem?\",\"answer\":\"QCCS uses a lightweight character n-gram, query-conditioned context selection gate to suppress unhelpful content and improve instruction-following accuracy compared with retrieval-only and reranking comparators.\"}]","INHIBITORY ATTENTION FOR CLINICAL LONGCONTEXT REASONING - Characterizing and Mitigating Lost-in-the-Middle Effects in EHR Processing | PDF",73]