[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"detail-sidebar-cat-0-en-105":3,"doc-seo-267458-105":59,"doc-detail-267458-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","sat-stepwise-adaptive-thinking-large-reasoning-models","SAT - Stepwise Adaptive Thinking - Large Reasoning Models","","Large Reasoning Models (LRMs) improve complex problem solving but often generate excessive “overthinking” with long Chains of Thought, reducing practicality in real-time and resource-limited settings. Existing token-efficiency approaches typically offer coarse control and may weaken logical fidelity or omit important fine-grained reasoning. Stepwise Adaptive Thinking (SAT) introduces step-level, difficulty-aware pruning using a finite-state reasoning controller and a lightweight Process Reward Model to compress easy steps while preserving depth for hard ones. Experiments across 9 LRMs and 7 benchmarks show up to 40% fewer reasoning tokens with maintained or improved accuracy.",{"@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/sat-stepwise-adaptive-thinking-large-reasoning-models/267458/",{"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/sat-stepwise-adaptive-thinking-large-reasoning-models/267458.png","ImageObject",300,407,{"name":92,"@type":93},"Chumphorn","Person",{"url":74,"name":95,"@type":96},"DocShare","Organization","application/pdf","2026-11-07","2026-09-14",true,{"@type":102,"interactionType":103,"userInteractionCount":105},"InteractionCounter",{"@type":104},"ViewAction",16,{"@type":107,"mainEntity":108},"FAQPage",[109,115,119],{"name":110,"@type":111,"acceptedAnswer":112},"What problem does SAT address in large reasoning models?","Question",{"text":113,"@type":114},"SAT targets the “overthinking” behavior where LRMs produce unnecessarily long Chains of Thought even for easy questions, wasting tokens and hindering real-time use.","Answer",{"name":116,"@type":111,"acceptedAnswer":117},"How does SAT control reasoning depth during inference?",{"text":118,"@type":114},"SAT formulates reasoning as a finite-state machine with thinking modes (SLOW, NORMAL, FAST, SKIP) and dynamically transitions between modes based on step-level difficulty.",{"name":120,"@type":111,"acceptedAnswer":121},"What role does the Process Reward Model (PRM) play in SAT?",{"text":122,"@type":114},"SAT uses a lightweight PRM to estimate a step-level difficulty score from confidence patterns and semantic information in the historical reasoning context, guiding whether to compress or elaborate each step.","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},267458,1789419382,{"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":56,"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":130,"read_time":144,"preview_page_count":4},2336475401981,"https://ap-avatar.wpscdn.com/avatar/22000c94efd8d5204d?x-image-process=image/resize,m_fixed,w_180,h_180&k=1786935347598174694","SAT: Balancing Reasoning Accuracy and Efficiency with Stepwise Adaptive  \nThinking  \nWeiyang Huang1 , Xuefeng Bai1 * , Kehai Chen1 , Xinyang Chen1 ,  \nYibin Chen2 , Weili Guan 1 , Min Zhang 1  \n1 Institute of Computing and Intelligence, Harbin Institute of Technology, Shenzhen, China  \n2 Huawei Technologies, China  \nAbstract  \nLarge Reasoning Models (LRMs) have revolutionized complex problem-solving, yet they exhibit a pervasive “overthinking”, generating unnecessarily long reasoning chains. While current solutions improve token efficiency, they often sacrifice fine-grained control or risk disrupting the logical integrity of the reasoning process. To address this, we introduce Stepwise Adaptive Thinking (SAT), a framework that performs step-level, difficulty-aware pruning while preserving the core reasoning structure.  \nSAT formulates reasoning as a Finite-State Machine (FSM) with distinct thinking modes (SLOW, NORMAL, FAST, SKIP) . It navigates these states dynamically using a lightweight Process Reward Model (PRM), compressing easy steps while preserving depth for hard ones.  \nExperiments across 9 LRMs and 7 benchmarks show that SAT achieves up to 40% reduction in reasoning tokens while generally maintaining or improving accuracy. Code is available at [https://github.com/byxw13/](https://github.com/byxw13/)[ ](https://github.com/byxw13/)SAT  Code.  \n1 Introduction  \nThe advent of large reasoning models (LRMs) such as OpenAI-O1 (OpenAI, 2025) and DeepSeekR1 (DeepSeek-AI et al., 2025) has marked a significant advance in tackling complex reasoning tasks (Wang et al., 2025b) . Nevertheless, these models are prone to a notable “overthinking” issue: they tend to produce excessively long Chains of Thought (COT) even for straightforward questions, creating a practical bottleneck in real-time or resource-constrained applications (Sui et al., 2025) . To mitigate this problem, numerous studies have sought to steer LLMs toward more token-efficient reasoning. Earlier work employs strategic prompting (Renze and Guven, 2024 ; Chen et al., 2024 ; Xu et al., 2025), question-driven routing (Liang  \n* Corresponding author  \nFigure 1: Baselines vs. SAT. COT spends redundant checks for easy steps, Early-Stopping halts after a highconfidence first-pass answer and fails, while SAT skips redundancy on easy steps but preserves verification on hard steps, achieving the correct answer efficiently.  \net al., 2025 ; Liu et al., 2025b), planning and budgeting (Han et al., 2025 ; Lin et al., 2025a) . However, these methods often regulate the reasoning process at a coarse-grained level, offering limited precise control over both the reasoning flow and its efficiency. A more recent and promising direction explores dynamic decoding mechanisms, such as early stopping of the COT sequence (Liu and Wang, 2025 ; Yang et al., 2025b), or suppressing reflective steps (Huang et al., 2026 ; Wang et al., 2025a) . While effective in shortening COT, these methods rely on a uniform policy across all steps, ignoring their varying difficulty and necessity. As shown in Figure 1, this strategy risks disrupting or cutting off essential steps, thereby compromising logical integrity and overall performance.  \nIn this work, we propose SAT (Stepwise Adaptive Thinking), a step-level, difficulty-aware pruning that eliminates redundancy while rigorously preserving the core logical structure essential for correct answers. As shown in Figure 1, departing from one-size-fits-all strategies, SAT is premised  \n43384  \nProceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages 43384–43402  \nJuly 2-7, 2026 ©2026 Association for Computational Linguistics  \non the observation that reasoning steps within a solution vary in difficulty (Saha et al., 2025) .1 Consequently, it dynamically navigates LLMs to employ a deeper thinking mode for challenging steps and a shallower, more efficient mode for simpler ones. Specifically, as illustrate","cbCaiaIQBZh0YyoD","https://ap.wps.com/l/cbCaiaIQBZh0YyoD","pdf",3015924,"English","# Introduction\n# Stepwise Adaptive Thinking (SAT)\n## Finite-State Machine thinking modes\n## Process Reward Model and difficulty scoring\n# Experiments and Results","[{\"question\":\"What problem does SAT address in large reasoning models?\",\"answer\":\"SAT targets the “overthinking” behavior where LRMs produce unnecessarily long Chains of Thought even for easy questions, wasting tokens and hindering real-time use.\"},{\"question\":\"How does SAT control reasoning depth during inference?\",\"answer\":\"SAT formulates reasoning as a finite-state machine with thinking modes (SLOW, NORMAL, FAST, SKIP) and dynamically transitions between modes based on step-level difficulty.\"},{\"question\":\"What role does the Process Reward Model (PRM) play in SAT?\",\"answer\":\"SAT uses a lightweight PRM to estimate a step-level difficulty score from confidence patterns and semantic information in the historical reasoning context, guiding whether to compress or elaborate each step.\"}]","SAT - Stepwise Adaptive Thinking - Large Reasoning Models | PDF",48]