[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"detail-sidebar-cat-0-en-105":3,"doc-seo-327482-105":59,"doc-detail-327482-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","driftlens-measuring-memory-induced-reasoning-drift-in-personalized-language-models","DRIFTLENS - Measuring Memory-Induced Reasoning Drift in Personalized Language Models","","Personalization changes not only what a language model says, but also the reasoning trajectory used to justify responses. The work studies whether persisted user-attribute memory reshapes reasoning on open-ended questions lacking a single ground-truth answer. It introduces DRIFTLENS, a ground-truth-free framework mapping each expressed reasoning step into a value category and measuring divergence between no-memory and injected-memory trajectories. Experiments across four LLMs and ten attribute categories show medium-to-large drift, and post-training mitigations reduce drift with model- and reward-dependent tradeoffs.",{"@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/driftlens-measuring-memory-induced-reasoning-drift-in-personalized-language-models/327482/",{"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/driftlens-measuring-memory-induced-reasoning-drift-in-personalized-language-models/327482.png","ImageObject",300,407,{"name":92,"@type":93},"Hazel","Person",{"url":74,"name":95,"@type":96},"DocShare","Organization","application/pdf","2026-11-08","2026-09-21",true,{"@type":102,"interactionType":103,"userInteractionCount":105},"InteractionCounter",{"@type":104},"ViewAction",17,{"@type":107,"mainEntity":108},"FAQPage",[109,115,119],{"name":110,"@type":111,"acceptedAnswer":112},"What problem does DRIFTLENS address in personalized language models?","Question",{"text":113,"@type":114},"It measures how persisted user-attribute memory can shift the reasoning trajectory used to justify responses, even when the final answers still sound fluent and plausible.","Answer",{"name":116,"@type":111,"acceptedAnswer":117},"Why can’t standard accuracy metrics reliably detect the drift studied here?",{"text":118,"@type":114},"The targeted tasks are open-ended and lack ground-truth answers, so accuracy cannot capture changes in the underlying chain of priorities and justifications.",{"name":120,"@type":111,"acceptedAnswer":121},"How is reasoning drift quantified in DRIFTLENS?",{"text":122,"@type":114},"DRIFTLENS uses a within-question no-memory trajectory as a baseline, maps expressed reasoning steps to a value ontology, and computes divergence between symbol sequences under injected user-attribute memory.","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},327482,1790455040,{"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":140,"language":141,"language_code":63,"site_id":62,"html_lang":63,"table_of_contents":142,"faqs":143,"seo_title":144,"seo_description":67,"update_tm":145,"read_time":146,"preview_page_count":4},137441390410,"https://ap-avatar.wpscdn.com/avatar/2000252f4ab5702993?_k=1776741390130283984","DRIFTLENS: Measuring Memory-Induced Reasoning Drift  \nin Personalized Language Models  \nXi Fang1 , Weijie Xu1 , Yingqiang Ge1 , Yuhui Xu1 ,  \nStephanie Eckman1 , Chandan K. Reddy1  \n1Amazon  \narXiv :2607 .02374v 1 [ cs .AI] 2 Jul 2026  \nAbstract  \nPersonalization changes what a model says toa user; we show that it can also change the reasoning trajectory used to justify the response. Modern LLMs personalize interactions by storing user attributes, preferences, and prior context, then injecting this information into future prompts. We study whether such memory reshapes reasoning on open-ended questions where no single ground-truth answer exists. To quantify this effect, we introduce DRIFTLENS, a ground-truth-free framework that maps each expressed reasoning step to a value category and measures divergence between a question’sno-memory trajectory and its trajectory under injected user-attribute memory. We first validate that DRIFTLENS distinguishes contentfree pragmatic noise from substantive reasoning changes. Across four LLMs and 10 userattribute categories, including age, occupation, and disability, user-attribute memory induces medium-to-large reasoning drift above each model’s pragmatic-noise floor, even when final answers remain fluent, on-topic, and plausible. We then evaluate GRPO- and DPObased post-training methods for reducing drift. Both reduce drift, but neither uniformly dominates; effects on downstream capability, helpfulness, and instruction following are modeland reward-dependent. These results suggest that memory-induced reasoning drift is a measurable and only partly mitigated failure mode of personalized language models.  \n Dataset: Datasets Repository  \n Code: Code Repository  \n1 Introduction  \nNearly 29% of real-world LLM interactions involve practical guidance, such as career decisions, interpersonal conflicts, and value-laden trade-offs, where no objectively correct answer exists (Chatterji et al., 2025 ; Shen et al., 2026) . Modern LLMs personalize these responses by extracting user at-  \nFigure 1: User memory reshapes reasoning. Question responses show different value trajectories (S1– S3) when a self-description is injected as context, shifting from protection/ethical responsibility to emotional depth/authentic connection.  \ntributes, persisting them across sessions, and conditioning subsequent conversations (OpenAI, 2024 ; Westhäußer et al., 2025 ; Jiang et al., 2025) . We show that this persisted context changes not just what a model says, but how it reasons. We call this phenomenon symbolic drift. Symbolic drift is dangerous precisely because it is invisible at the answer level: a model may still sound reasonable while relying on a different chain of priorities, trade-offs, and justifications. Because these tasks lack groundtruth answers, standard accuracy metrics cannot detect the shift; reliability must instead be measured through reasoning stability, that is, whether the model’s decision process changes when the context around the question changes.  \nSymbolic drift is expected when context is relevant, since adapting to user needs is the intended behavior of memory-equipped systems. But when questions are independent of user attributes, any drift caused by those attributes signals unwanted sensitivity rather than appropriate adaptation. For example, when asked about a workplace dispute, a model should not shift from legal-risk reasoning to emotional-validation reasoning simply because the stored profile indicates the user is young, disabled, low-income, or in a stigmatized occupation, unless the question itself makes that attribute relevant. Au-  \ndits reveal that systems extract and store memories without explicit user action and that these memories frequently encode psychological traits (Dash et al., 2026), leaving users with limited ability toopt out of the conditioning that drives drift. We propose DRIFTLENS, which measures per-instance reasoning stability under memory perturbations in single ","cbCaidJObzeyebZR","https://ap.wps.com/l/cbCaidJObzeyebZR","pdf",4969205,39,"English","# Abstract\n# 1 Introduction\n# 2 The DRIFTLENS Framework\n## 2.1 Data Curation\n## 2.2 Value Ontology\n## 2.3 User-Memory Perturbations\n## 2.4 Drift Metrics\n## 2.5 Statistical Model","[{\"question\":\"What problem does DRIFTLENS address in personalized language models?\",\"answer\":\"It measures how persisted user-attribute memory can shift the reasoning trajectory used to justify responses, even when the final answers still sound fluent and plausible.\"},{\"question\":\"Why can’t standard accuracy metrics reliably detect the drift studied here?\",\"answer\":\"The targeted tasks are open-ended and lack ground-truth answers, so accuracy cannot capture changes in the underlying chain of priorities and justifications.\"},{\"question\":\"How is reasoning drift quantified in DRIFTLENS?\",\"answer\":\"DRIFTLENS uses a within-question no-memory trajectory as a baseline, maps expressed reasoning steps to a value ontology, and computes divergence between symbol sequences under injected user-attribute memory.\"}]","DRIFTLENS - Measuring Memory-Induced Reasoning Drift in Personalized Language Models | PDF",1789977302,98]