[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-160262-en":3,"doc-seo-160262-105":30,"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":4,"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":27,"seo_description":14,"update_tm":28,"read_time":29},160262,2336475104042,"Tawan","https://ap-avatar.wpscdn.com/avatar/22000c4c32af1715be0?x-image-process=image/resize,m_fixed,w_180,h_180&k=1786537525561427321",8,"Research & Report","PCA-guided Activation Scaling for Monotonic Bidirectional Control over LLM Sycophancy","Large language models exhibit sycophancy, the tendency to agree with user beliefs regardless of factual accuracy, which can intensify misconceptions and reduce trust. Eliminating sycophancy completely risks over-correction that undermines valid user opinions, so effective control must be both bidirectional and monotonic with predictable gradual effects. Existing techniques do not guarantee monotonic steering across models and datasets. This work proposes PCA-guided Activation Scaling (PAS), decomposing activations into a PCA-identified sycophancy-honesty subspace and an orthogonal residual, then applying distinct scaling exponents for controlled shifts. Experiments on three LLMs and three datasets show strong monotonicity (Spearman ρ = +0.92) and larger average behavioral shifts than baselines, with ablations confirming essential components.","arXiv :2608 . 16650v1 [ cs .CL] 17 Aug 2026  \nPCA-guided Activation Scaling for Monotonic Bidirectional Control over LLM Sycophancy  \nZheng Chen1, Zhaoxin Feng2, Yip Tin Po1, Jianfei Ma2, Emmanuele Chersoni2, Bo Li1  \n1The Hong Kong University of Science and Technology  \n2The Hong Kong Polytechnic University  \n{zchenin, [tpyip](tpyip}@connect.ust.hk)[}](tpyip}@connect.ust.hk)[@connect.ust.hk](tpyip}@connect.ust.hk) [bli@cse.ust.hk](bli@cse.ust.hk)[ ](bli@cse.ust.hk){zhaoxinbetty.feng, [jianfei-mark.ma](jianfei-mark.ma}@connect.polyu.hk)[}](jianfei-mark.ma}@connect.polyu.hk)[@connect.polyu.hk](jianfei-mark.ma}@connect.polyu.hk)[ ](jianfei-mark.ma}@connect.polyu.hk)[emmanuele.chersoni@polyu.edu.hk](emmanuele.chersoni@polyu.edu.hk)  \nAbstract  \nLarge language models (LLMs) exhibit sycophancy, a tendency to agree with user beliefs regardless of factual accuracy. This can reinforce misconceptions, but eliminating it entirely risks over-correction against valid opinions. Effective control must therefore both reduce and increase sycophancy with predictable and gradual effect. Yet, existing methods fail to ensure a bidirectional and monotonic relationship between steering strength and behavioral outcome across models and datasets. We introduce PCA-guided Activation Scaling (PAS), an activation steering framework that decomposes residual stream activations into a PCA-identified sycophancy-honesty subspace and an orthogonal residual, then applies distinct scaling exponents to achieve monotonic, bidirectional control. Across three LLMs and three datasets, PAS achieves strong monotonicity (Spearman ρ = +0.92) and an average shift of 15.4% per direction, compared with 8.7% for the baselines. Ablation studies confirm that the decomposition, asymmetric exponents, and layer selection are each essential for maintaining monotonic control. The data and code are available at [https://github.com/Bellafc/PCS](https://github.com/Bellafc/PCS).  \n1 Introduction  \nLarge language models (LLMs) demonstrate remarkable capabilities across diverse tasks, yet they exhibit a persistent tendency known as sycophancy: aligning responses with user beliefs regardless of factual accuracy (Perez et al., 2023; Sharma et al., 2024) . Sycophantic behavior can reinforce users’ factual misconceptions instead of providing accurate corrections (Chenet al., 2025), causing misinformation and reduced trust (Wei et al., 2024; Perez et al., 2023) . However, naively eliminating sycophancy carries risks (Figure 1): over-correction may cause models to become unnecessarily adversarial, refusing to acknowledge users’ valid opinions even when agreement is warranted (Seitz, 2024; Clegg, 2025) .  \nThis challenge necessitates control mechanisms that can both reduce and increase sycophantic behavior depending on the deployment context (Vennemeyer et al., 2026), with predictable and gradual effects as the control parameter varies. That is, effective sycophancy control must be both bidirectional and monotonic: adjusting a single parameter should reliably shift behavior along the sycophancy-honesty spectrum.  \nExisting approaches to sycophancy mitigation fall short of this requirement. Training-time interventions, such as fine-tuning on synthetic data (Wei et al., 2024) or reward modeling (Ram et al., 2024), require expensive dataset curation and model retraining for each desired behavior point. Prompt engineering techniques (Turpin et al., 2023; Lyu et al., 2025) can influence outputs but lack continuous control and vary widely across models and tasks.  \nActivation steering has emerged as a promising alternative that addresses these limitations by intervening on internal representations at inference time (Turner et al., 2024; Li et al., 2024;  \nZou et al., 2023) . Unlike training methods that require expensive large-scale dataset curation and model retraining, steering operates with minimal data; unlike prompting, it enables uniform control via internal interventions rather than per-sample","cbCainfI4grrfJab","https://ap.wps.com/l/cbCainfI4grrfJab","pdf",1573346,1,30,"English","en",105,"# Abstract\n# Introduction\n## Problem: sycophancy and its risks\n## Requirements: bidirectional and monotonic control\n## Limitations of existing approaches\n## Proposed method: PCA-guided Activation Scaling (PAS)\n# Contributions","[{\"question\":\"What problem does PAS address in large language models?\",\"answer\":\"PAS addresses sycophancy in LLMs, where responses align with user beliefs even when accuracy is not ensured. The goal is to control this tendency rather than remove it blindly.\"},{\"question\":\"Why does the document emphasize bidirectional and monotonic control?\",\"answer\":\"The work argues that effective control should both reduce and increase sycophancy depending on context, and changes should follow predictable, gradual behavior as the steering parameter varies.\"},{\"question\":\"How does PCA-guided Activation Scaling (PAS) achieve the control mechanism?\",\"answer\":\"PAS decomposes residual stream activations into a PCA-identified sycophancy-honesty subspace and an orthogonal residual, then applies different scaling exponents to each component to enforce monotonic, bidirectional effects.\"}]","PCA-guided Activation Scaling for Monotonic Bidirectional Control over LLM Sycophancy | PDF",1788052987,76,{"code":4,"msg":31,"data":32},"ok",{"site_id":24,"language":23,"slug":33,"title":13,"keywords":34,"description":14,"schema_data":35,"social_meta":86,"head_meta":88,"extra_data":90,"updated_unix":28},"pca-guided-activation-scaling-for-monotonic-bidirectional-control-over-llm-sycophancy","",{"@graph":36,"@context":85},[37,54,68],{"@type":38,"itemListElement":39},"BreadcrumbList",[40,44,48,51],{"item":41,"name":42,"@type":43,"position":20},"https://docshare.wps.com","Home","ListItem",{"item":45,"name":46,"@type":43,"position":47},"https://docshare.wps.com/document/","Document",2,{"item":49,"name":12,"@type":43,"position":50},"https://docshare.wps.com/document/research-report/",3,{"item":52,"name":13,"@type":43,"position":53},"https://docshare.wps.com/document/pca-guided-activation-scaling-for-monotonic-bidirectional-control-over-llm-sycophancy/160262/",4,{"url":52,"name":13,"@type":55,"author":56,"headline":13,"publisher":58,"fileFormat":61,"inLanguage":23,"description":14,"dateModified":62,"datePublished":62,"encodingFormat":61,"isAccessibleForFree":63,"interactionStatistic":64},"DigitalDocument",{"name":9,"@type":57},"Person",{"url":41,"name":59,"@type":60},"DocShare","Organization","application/pdf","2026-08-30",true,{"@type":65,"interactionType":66,"userInteractionCount":4},"InteractionCounter",{"@type":67},"ViewAction",{"@type":69,"mainEntity":70},"FAQPage",[71,77,81],{"name":72,"@type":73,"acceptedAnswer":74},"What problem does PAS address in large language models?","Question",{"text":75,"@type":76},"PAS addresses sycophancy in LLMs, where responses align with user beliefs even when accuracy is not ensured. The goal is to control this tendency rather than remove it blindly.","Answer",{"name":78,"@type":73,"acceptedAnswer":79},"Why does the document emphasize bidirectional and monotonic control?",{"text":80,"@type":76},"The work argues that effective control should both reduce and increase sycophancy depending on context, and changes should follow predictable, gradual behavior as the steering parameter varies.",{"name":82,"@type":73,"acceptedAnswer":83},"How does PCA-guided Activation Scaling (PAS) achieve the control mechanism?",{"text":84,"@type":76},"PAS decomposes residual stream activations into a PCA-identified sycophancy-honesty subspace and an orthogonal residual, then applies different scaling exponents to each component to enforce monotonic, bidirectional effects.","https://schema.org",{"og:url":52,"og:type":87,"og:title":13,"og:site_name":59,"og:description":14},"article",{"robots":89,"canonical":52},"index,follow",{"doc_id":7,"site_id":24},{"code":4,"msg":5,"data":92},[93,97,101,105,110,115,120,122,127,130,134],{"id":20,"doc_module":4,"doc_module_name":46,"category_name":94,"show_sort_weight":95,"slug":96},"Story & Novel",90,"story-novel",{"id":47,"doc_module":4,"doc_module_name":46,"category_name":98,"show_sort_weight":99,"slug":100},"Literature",80,"literature",{"id":53,"doc_module":4,"doc_module_name":46,"category_name":102,"show_sort_weight":103,"slug":104},"Exam",70,"exam",{"id":106,"doc_module":4,"doc_module_name":46,"category_name":107,"show_sort_weight":108,"slug":109},5,"Comic",60,"comic",{"id":111,"doc_module":4,"doc_module_name":46,"category_name":112,"show_sort_weight":113,"slug":114},6,"Technology",50,"technology",{"id":116,"doc_module":4,"doc_module_name":46,"category_name":117,"show_sort_weight":118,"slug":119},7,"Healthcare",40,"healthcare",{"id":11,"doc_module":4,"doc_module_name":46,"category_name":12,"show_sort_weight":21,"slug":121},"research-report",{"id":123,"doc_module":4,"doc_module_name":46,"category_name":124,"show_sort_weight":125,"slug":126},9,"Religion & Spirituality",20,"religion-spirituality",{"id":125,"doc_module":4,"doc_module_name":46,"category_name":128,"show_sort_weight":125,"slug":129},"World Cup","world-cup",{"id":131,"doc_module":4,"doc_module_name":46,"category_name":132,"show_sort_weight":131,"slug":133},10,"Lifestyle","lifestyle",{"id":135,"doc_module":4,"doc_module_name":46,"category_name":136,"show_sort_weight":106,"slug":137},19,"General","general"]