[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-seo-573890-105":3,"detail-sidebar-cat-1-en-105":74,"doc-detail-573890-en":120},{"code":4,"msg":5,"data":6},0,"ok",{"site_id":7,"language":8,"slug":9,"title":10,"keywords":11,"description":12,"schema_data":13,"social_meta":67,"head_meta":69,"extra_data":71,"updated_unix":73},105,"en","distilling-many-shot-in-context-learning-into-a-cheat-sheet","Distilling Many-Shot In-Context Learning into a Cheat Sheet","","Recent advances in large language models enable many-shot in-context learning (ICL), yet it becomes computationally expensive because longer inputs require more tokens. The paper introduces cheat-sheet ICL, which compresses knowledge from many-shot demonstrations into a concise textual cheat sheet used as inference-time context. Experiments on demanding reasoning tasks show cheat-sheet ICL matches or improves over many-shot ICL using far fewer tokens and also reaches retrieval-based ICL performance without test-time retrieval, supporting practical deployment for downstream tasks.",{"@graph":14,"@context":66},[15,34,49],{"@type":16,"itemListElement":17},"BreadcrumbList",[18,23,27,31],{"item":19,"name":20,"@type":21,"position":22},"https://docshare.wps.com","Home","ListItem",1,{"item":24,"name":25,"@type":21,"position":26},"https://docshare.wps.com/template/","Template",2,{"item":28,"name":29,"@type":21,"position":30},"https://docshare.wps.com/template/presentations/","Presentations",3,{"item":32,"name":10,"@type":21,"position":33},"https://docshare.wps.com/template/distilling-many-shot-in-context-learning-into-a-cheat-sheet/573890/",4,{"url":32,"name":10,"@type":35,"author":36,"headline":10,"publisher":39,"fileFormat":42,"inLanguage":8,"description":12,"dateModified":43,"datePublished":43,"encodingFormat":42,"isAccessibleForFree":44,"interactionStatistic":45},"DigitalDocument",{"name":37,"@type":38},"dabgoudw","Person",{"url":19,"name":40,"@type":41},"DocShare","Organization","application/pdf","2026-10-14",true,{"@type":46,"interactionType":47,"userInteractionCount":4},"InteractionCounter",{"@type":48},"ViewAction",{"@type":50,"mainEntity":51},"FAQPage",[52,58,62],{"name":53,"@type":54,"acceptedAnswer":55},"What problem does cheat-sheet ICL address?","Question",{"text":56,"@type":57},"Cheat-sheet ICL targets the high computational cost of many-shot in-context learning caused by longer input sequences and more tokens during inference.","Answer",{"name":59,"@type":54,"acceptedAnswer":60},"How does cheat-sheet ICL differ from many-shot ICL at inference time?",{"text":61,"@type":57},"It distills information from many-shot demonstrations into a compact textual cheat sheet that serves as context at inference time, reducing the need to process many demonstrations each time.",{"name":63,"@type":54,"acceptedAnswer":64},"Does cheat-sheet ICL require retrieval at test time?",{"text":65,"@type":57},"No. Experiments indicate cheat-sheet ICL matches the performance of retrieval-based ICL without requiring any test-time retrieval step.","https://schema.org",{"og:url":32,"og:type":68,"og:title":10,"og:site_name":40,"og:description":12},"article",{"robots":70,"canonical":32},"index,follow",{"doc_id":72,"site_id":7},573890,1791950517,{"code":4,"msg":75,"data":76},"success",[77,81,86,91,96,101,106,111,116],{"id":78,"doc_module":22,"doc_module_name":25,"category_name":29,"show_sort_weight":79,"slug":80},11,90,"presentations",{"id":82,"doc_module":22,"doc_module_name":25,"category_name":83,"show_sort_weight":84,"slug":85},12,"Resumes",80,"resumes",{"id":87,"doc_module":22,"doc_module_name":25,"category_name":88,"show_sort_weight":89,"slug":90},14,"Invoices",70,"invoices",{"id":92,"doc_module":22,"doc_module_name":25,"category_name":93,"show_sort_weight":94,"slug":95},15,"Posters",60,"posters",{"id":97,"doc_module":22,"doc_module_name":25,"category_name":98,"show_sort_weight":99,"slug":100},16,"Social Media",50,"social-media",{"id":102,"doc_module":22,"doc_module_name":25,"category_name":103,"show_sort_weight":104,"slug":105},17,"Forms",40,"forms",{"id":107,"doc_module":22,"doc_module_name":25,"category_name":108,"show_sort_weight":109,"slug":110},18,"Letters",30,"letters",{"id":112,"doc_module":22,"doc_module_name":25,"category_name":113,"show_sort_weight":114,"slug":115},21,"Paper Templates",5,"papers-templates",{"id":117,"doc_module":22,"doc_module_name":25,"category_name":118,"show_sort_weight":4,"slug":119},158,"General","general-158",{"code":4,"msg":75,"data":121},{"doc_id":72,"user_id":122,"nickname":37,"user_avatar":123,"doc_module":22,"category_id":78,"category_name":29,"doc_title":10,"doc_description":12,"doc_content":124,"file_id":125,"file_url":126,"file_type":127,"file_size":128,"view_count":4,"is_deleted":4,"is_public":22,"is_downloadable":22,"audit_status":22,"page_count":112,"language":129,"language_code":8,"site_id":7,"html_lang":8,"table_of_contents":130,"faqs":131,"seo_title":132,"seo_description":12,"update_tm":73,"read_time":133},3985747860484,"https://ap-avatar.wpscdn.com/davatar_276721f389ce27ea32af1340a28f341c","Distilling Many-Shot In-Context Learning into a Cheat Sheet  \nUkyo Honda Soichiro Murakami Peinan Zhang  \nCyberAgent, Tokyo, Japan  \n{honda_ukyo,murakami_soichiro,[zhang_peinan}@cyberagent.co.jp](zhang_peinan}@cyberagent.co.jp)  \narXiv :2509 .20820v1 [ cs .CL] 25 Sep 2025  \nAbstract  \nRecent advances in large language models (LLMs) enable effective in-context learning (ICL) with many-shot examples, but at the cost of high computational demand due to longer input tokens. To address this, we propose cheat-sheet ICL, which distills the information from many-shot ICL into a concise textual summary (cheat sheet) used as the context at inference time. Experiments on challenging reasoning tasks show that cheat-sheet ICL achieves comparable or better performance than many-shot ICL with far fewer tokens, and matches retrieval-based ICL without requiring test-time retrieval. These findings demonstrate that cheat-sheet ICL is a practical alternative for leveraging LLMs in downstream tasks.1  \n1 Introduction  \nIn-context learning (ICL; Brown et al., 2020) has emerged as a novel paradigm for leveraging large language models (LLMs) on downstream tasks (Dong et al., 2024) . Unlike the predominant finetuning approach, ICL does not involve updating the model parameters. Instead, it provides a few taskspecific examples, known as demonstrations, anda test input together, enabling LLMs to infer based on this context. Due to context window limitations, ICL has been typically used in few-shot settings.  \nBuilding on the extended context windows afforded by recent advancements in LLMs, Agarwalet al. (2024) and Bertsch et al. (2025) have demonstrated the superior performance of many-shot ICL, wherein a larger number of demonstrations are provided, over the conventional few-shot setup. Although many-shot ICL requires a larger number of demonstrations, it retains key advantages over fine-tuning: it is training-free and can be applied to state-of-the-art proprietary models, which often limit or do not support fine-tuning. Nevertheless,  \n1The code is publicly available at [https://github.com/](https://github.com/)[ ](https://github.com/)CyberAgentAILab/cheat-sheet-icl.  \nFigure 1: Overview of many-shot, retrieval-based, and our cheat-sheet ICL. Many-shot ICL processes numerous demonstrations at every inference; retrieval-based ICL involves storing and searching them. Cheat-sheet ICL uses a compact cheat sheet distilled from demonstrations, which only needs to be created once.  \nthis paradigm introduces the increased computational costs associated with providing substantially longer input contexts.  \nA common approach to efficient ICL, given a large pool of demonstrations, is demonstration retrieval, in which demonstrations are selected based on their similarity to test inputs (Liu et al., 2022 ; Bertsch et al., 2025) . While effective, this method requires a retrieval operation for each inference.  \nIn this study, we explore an alternative to existing approaches. Our key idea is that LLMs, with their advanced language understanding, can extract the knowledge needed for a task as a textual summary. Rather than requiring LLMs to infer hidden patterns from demonstrations at each inference, we propose distilling many-shot ICL knowledge into an explicit textual format. Figure 1 presents an  \noverview of our method. We term our approach cheat-sheet ICL, by analogy to how students summarize key points on a single sheet for exams.  \nExperimental results show that this remarkably simple method achieves performance comparable to, or even surpassing, many-shot ICL on challenging reasoning tasks. Furthermore, cheat sheets are interpretable and allow easy intervention for further improvements. Moreover, cheat-sheet ICL matches demonstration retrieval in performance, indicating its viability as an alternative. We consider this alternative paradigm a promising direction, and anticipate further improvements as LLMs’ language understanding advances.  \n2 Preliminaries  \n","cbCaitMyMKjYEevU","https://ap.wps.com/l/cbCaitMyMKjYEevU","pdf",755271,"English","# Abstract\n# Introduction\n# Preliminaries\n## ICL and Many-Shot ICL\n## Improved ICL Baseline\n# Method: Cheat-Sheet ICL","[{\"question\":\"What problem does cheat-sheet ICL address?\",\"answer\":\"Cheat-sheet ICL targets the high computational cost of many-shot in-context learning caused by longer input sequences and more tokens during inference.\"},{\"question\":\"How does cheat-sheet ICL differ from many-shot ICL at inference time?\",\"answer\":\"It distills information from many-shot demonstrations into a compact textual cheat sheet that serves as context at inference time, reducing the need to process many demonstrations each time.\"},{\"question\":\"Does cheat-sheet ICL require retrieval at test time?\",\"answer\":\"No. Experiments indicate cheat-sheet ICL matches the performance of retrieval-based ICL without requiring any test-time retrieval step.\"}]","Distilling Many-Shot In-Context Learning into a Cheat Sheet | PDF",7]