[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"detail-sidebar-cat-0-en-105":3,"doc-seo-273869-105":59,"doc-detail-273869-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","memory-augmentation-unlocks-efficient-chain-of-thought-reasoning","Memory Augmentation Unlocks Efficient Chain-of-Thought Reasoning","","Large language models often rely on Chain-of-Thought (CoT) reasoning, but verbose reasoning traces create substantial inference overhead. Compression methods can reduce generation cost, yet overly aggressive compression may harm logical coherence and performance. The work formalizes this trade-off via the Context-Generation Substitution Law, then introduces Memory-Augmented Compression, a training-free framework that builds reusable reasoning memories from past traces and retrieves them as prefill-side scaffolds. Experiments show accuracy gains and 1.14–1.49× latency improvements.",{"@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/memory-augmentation-unlocks-efficient-chain-of-thought-reasoning/273869/",{"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/memory-augmentation-unlocks-efficient-chain-of-thought-reasoning/273869.png","ImageObject",300,407,{"name":92,"@type":93},"Ophelia","Person",{"url":74,"name":95,"@type":96},"DocShare","Organization","application/pdf","2026-10-29","2026-09-15",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},"Why does verbose Chain-of-Thought reasoning increase cost in LLMs?","Question",{"text":113,"@type":114},"CoT traces must be generated token by token during autoregressive decoding, which raises decoding latency, token cost, and serving overhead.","Answer",{"name":116,"@type":111,"acceptedAnswer":117},"What is the Context-Generation Substitution Law?",{"text":118,"@type":114},"It characterizes the trade-off between using reasoning context in the prefill stage versus generating reasoning during decoding, treating explicit reasoning context as a partial substitute for decode-time generation.",{"name":120,"@type":111,"acceptedAnswer":121},"How does Memory-Augmented Compression improve compressed reasoning?",{"text":122,"@type":114},"It retrieves relevant abstract reasoning memories built from historical traces and injects them into prompts as prefill-side scaffolds, compensating for information lost during compression.","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},273869,1789473084,{"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},7971461741311,"https://ap-avatar.wpscdn.com/avatar/74000253aff267980c6?x-image-process=image/resize,m_fixed,w_180,h_180&k=1779345379180704826","Memory Augmentation Unlocks Efficient Chain-of-Thought Reasoning  \nSimeng Zhang 1 ,2 ,∗ , Yilong Chen 1 ,2 ,∗ , Wenyuan Zhang4  \nZhenyu Zhang3 , Yao Chen 1 ,2 , Junyuan Shang3 , Tingwen Liu 1 ,2†  \n1 Institute of Information Engineering, Chinese Academy of Sciences  \n2 School of Cyber Security, University of Chinese Academy of Sciences  \n3 Baidu Inc. 4 Tencent Inc.  \n{zhangsimeng, chenyilong, [liutingwen}@iie.ac.cn](liutingwen}@iie.ac.cn)  \narXiv : 2608 .2 1265v2 [cs .CL] 26 Aug 2026  \nAbstract  \nLarge language models often rely on Chainof-Thought (CoT) reasoning to solve complex tasks, but verbose reasoning traces introduce substantial inference overhead. CoT compression shortens generation, yet aggressive compression may disrupt logical coherence and degrade performance. We formalize this trade-off as the Context-Generation Substitution Law, where explicit reasoning context substitutes for part of decode-time generation. Based on this principle, we propose Memory-Augmented Compression, a trainingfree framework that constructs reusable reasoning memories from historical traces and retrieves them as prefill-side scaffolds. Rather than using raw demonstrations, these memories summarize reusable reasoning patterns, key constraints, and critical operations to compensate for information lost during compression. Experiments show that Memory consistently improves prompt-based Chain-of-Draft (CoD) compression across mathematical reasoning, complex reasoning, and science question answering tasks, yielding accuracy gains of 21.4, 28.0, 29.5, and 6.61 points over CoDon GSM8K, MATH, BBH, and MMLU-Sci, while achieving a 1.14–1.49× latency speedup over standard CoT. Memory is also compatible with token-level, reasoning-trace-level, and inference-state compression mechanisms. Further analyzes show that the gains come from relevant reasoning memories rather than simply increasing context length.  \n1 Introduction  \nLarge Language Models (LLMs) exhibit strong generalization capabilities, enabling them to perform complex reasoning tasks. Chain-of-Thought  \n∗ Equal contribution. † Corresponding author.  \nAn initial version of this work was completed in January 2026 .  \nFigure 1: Illustration of memory-augmented compression. Explicit memory shifts part of the reasoning support to the prefill stage, improving short-CoT accuracy while preserving generation efficiency.  \n(CoT) prompting (Wei et al., 2022) improves reasoning by explicitly generating intermediate steps, while recent reasoning-oriented models such as OpenAI o-series models (Jaech et al., 2024) and DeepSeek-R1 (Guo et al., 2025) further show that scaling test-time computation and producing longer reasoning traces can substantially improve performance on difficult benchmarks. However, this longthinking paradigm comes with a significant cost: verbose reasoning traces must be generated token by token in an autoregressive manner, leading to increased decoding latency, token cost, and serving overhead (Kwon et al., 2023 ; Chen et al., 2026) .  \nTo mitigate the overhead of long reasoning traces, existing methods compress reasoning at different levels, including semantic-level shortening, token-level pruning or skipping, and inference-state compression (Xu et al., 2025 ; Xia et al., 2025 ; Songet al., 2026) . However, these methods mainly reduce reasoning information produced or retained during generation. As compression becomes more aggressive, they may discard useful subgoals, constraints, or logical dependencies, causing accuracy degradation. This exposes a key limitation of generation-side compression: shorter reasoning traces do not necessarily preserve the support needed for problem solving. How, then, can we  \npreserve such support without returning to long autoregressive generation?  \nHow can we leverage the parallelism of context scaling to substitute the serial overhead of lengthy generation?  \nMotivated by this question, we conduct a preliminary study on the interplay between input ","cbCaibCiQEfdPMtV","https://ap.wps.com/l/cbCaibCiQEfdPMtV","pdf",1352542,"English","# Abstract\n# 1 Introduction\n## Contributions\n# 2 Preliminaries\n## 2.1 Standard Chain-of-T","[{\"question\":\"Why does verbose Chain-of-Thought reasoning increase cost in LLMs?\",\"answer\":\"CoT traces must be generated token by token during autoregressive decoding, which raises decoding latency, token cost, and serving overhead.\"},{\"question\":\"What is the Context-Generation Substitution Law?\",\"answer\":\"It characterizes the trade-off between using reasoning context in the prefill stage versus generating reasoning during decoding, treating explicit reasoning context as a partial substitute for decode-time generation.\"},{\"question\":\"How does Memory-Augmented Compression improve compressed reasoning?\",\"answer\":\"It retrieves relevant abstract reasoning memories built from historical traces and injects them into prompts as prefill-side scaffolds, compensating for information lost during compression.\"}]","Memory Augmentation Unlocks Efficient Chain-of-Thought Reasoning | PDF",48]