[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"detail-sidebar-cat-1-en-105":3,"doc-seo-572637-105":53,"doc-detail-572637-en":126},{"code":4,"msg":5,"data":6},0,"success",[7,14,19,24,29,34,39,44,49],{"id":8,"doc_module":9,"doc_module_name":10,"category_name":11,"show_sort_weight":12,"slug":13},11,1,"Template","Presentations",90,"presentations",{"id":15,"doc_module":9,"doc_module_name":10,"category_name":16,"show_sort_weight":17,"slug":18},12,"Resumes",80,"resumes",{"id":20,"doc_module":9,"doc_module_name":10,"category_name":21,"show_sort_weight":22,"slug":23},14,"Invoices",70,"invoices",{"id":25,"doc_module":9,"doc_module_name":10,"category_name":26,"show_sort_weight":27,"slug":28},15,"Posters",60,"posters",{"id":30,"doc_module":9,"doc_module_name":10,"category_name":31,"show_sort_weight":32,"slug":33},16,"Social Media",50,"social-media",{"id":35,"doc_module":9,"doc_module_name":10,"category_name":36,"show_sort_weight":37,"slug":38},17,"Forms",40,"forms",{"id":40,"doc_module":9,"doc_module_name":10,"category_name":41,"show_sort_weight":42,"slug":43},18,"Letters",30,"letters",{"id":45,"doc_module":9,"doc_module_name":10,"category_name":46,"show_sort_weight":47,"slug":48},21,"Paper Templates",5,"papers-templates",{"id":50,"doc_module":9,"doc_module_name":10,"category_name":51,"show_sort_weight":4,"slug":52},158,"General","general-158",{"code":4,"msg":54,"data":55},"ok",{"site_id":56,"language":57,"slug":58,"title":59,"keywords":60,"description":61,"schema_data":62,"social_meta":119,"head_meta":121,"extra_data":123,"updated_unix":125},105,"en","dynamic-cheatsheet-dynamic-cheatsheet-test-time-learning-with-adaptive-memory","Dynamic Cheatsheet - Dynamic Cheatsheet - Test-Time Learning with Adaptive Memory","","Dynamic Cheatsheet (DC) is a lightweight test-time learning framework that equips black-box language models with a persistent, evolving memory. Instead of treating each query independently, DC dynamically curates a compact library of reusable strategies, solution sketches, and code snippets, storing lessons from successes and revising or pruning from errors at inference time. Experiments show large accuracy gains across math and knowledge-heavy benchmarks without needing ground-truth labels, human feedback, fine-tuning, or parameter updates.",{"@graph":63,"@context":118},[64,80,101],{"@type":65,"itemListElement":66},"BreadcrumbList",[67,71,74,77],{"item":68,"name":69,"@type":70,"position":9},"https://docshare.wps.com","Home","ListItem",{"item":72,"name":10,"@type":70,"position":73},"https://docshare.wps.com/template/",2,{"item":75,"name":11,"@type":70,"position":76},"https://docshare.wps.com/template/presentations/",3,{"item":78,"name":59,"@type":70,"position":79},"https://docshare.wps.com/template/dynamic-cheatsheet-dynamic-cheatsheet-test-time-learning-with-adaptive-memory/572637/",4,{"url":78,"name":59,"@type":81,"image":82,"author":87,"headline":59,"publisher":90,"fileFormat":93,"inLanguage":57,"description":61,"dateModified":94,"datePublished":95,"encodingFormat":93,"isAccessibleForFree":96,"interactionStatistic":97},"DigitalDocument",{"url":83,"@type":84,"width":85,"height":86},"https://docshare.wps.com/thumbnails/dynamic-cheatsheet-dynamic-cheatsheet-test-time-learning-with-adaptive-memory/572637.png","ImageObject",442,249,{"name":88,"@type":89},"gewufjwa","Person",{"url":68,"name":91,"@type":92},"DocShare","Organization","application/pdf","2026-10-20","2026-10-13",true,{"@type":98,"interactionType":99,"userInteractionCount":79},"InteractionCounter",{"@type":100},"ViewAction",{"@type":102,"mainEntity":103},"FAQPage",[104,110,114],{"name":105,"@type":106,"acceptedAnswer":107},"What problem does Dynamic Cheatsheet address for language models?","Question",{"text":108,"@type":109},"It addresses the limitation that deployed language models typically have no persistent memory of prior questions, successes, or mistakes during inference, causing repeated re-derivation of the same insights and repeated errors.","Answer",{"name":111,"@type":106,"acceptedAnswer":112},"How does DC work during inference time?",{"text":113,"@type":109},"DC first consults an external memory to retrieve relevant stored lessons, generates a solution by combining retrieved insights with internal reasoning, and then updates the memory via a curation step based on whether the approach is correct or an error occurs.",{"name":115,"@type":106,"acceptedAnswer":116},"Does DC require fine-tuning or ground-truth labels?",{"text":117,"@type":109},"No. DC is designed to improve models at inference time without gradient-based parameter updates and without needing explicit ground-truth labels or human feedback.","https://schema.org",{"og:url":78,"og:type":120,"og:title":59,"og:site_name":91,"og:description":61},"article",{"robots":122,"canonical":78},"index,follow",{"doc_id":124,"site_id":56},572637,1792360101,{"code":4,"msg":5,"data":127},{"doc_id":124,"user_id":128,"nickname":88,"user_avatar":129,"doc_module":9,"category_id":8,"category_name":11,"doc_title":59,"doc_description":61,"doc_content":130,"file_id":131,"file_url":132,"file_type":133,"file_size":134,"view_count":79,"is_deleted":4,"is_public":9,"is_downloadable":9,"audit_status":9,"page_count":135,"language":136,"language_code":57,"site_id":56,"html_lang":57,"table_of_contents":137,"faqs":138,"seo_title":139,"seo_description":61,"update_tm":140,"read_time":141},3985747882900,"https://ap-avatar.wpscdn.com/davatar_9964176cb1d06d4a9deccf72a44ae3dc","arXiv :2504 .07952v1 [ cs .LG] 10 Apr 2025  \nDynamic Cheatsheet: Test-Time Learning with Adaptive Memory  \nMirac Suzgun 1 Mert Yuksekgonul 1 Federico Bianchi 2 Dan Jurafsky 1 James Zou 1 2  \nAbstract  \nDespite their impressive performance on complex tasks, current language models (LMs) typically operate ina vacuum: Each input query is processed separately, without retaining insights from previous attempts. Here, we present Dynamic Cheatsheet (DC), a lightweight framework that endows a black-box LM with a persistent, evolving memory. Rather than repeatedly re-discovering or re-committing the same solutions and mistakes, DC enables models to store and reuse accumulated strategies, code snippets, and general problem-solving insights at inference time. This test-time learning enhances performance substantially across a range of tasks without needing explicit ground-truth labels or human feedback. Leveraging DC, Claude 3.5 Sonnet’s accuracy more than doubledon AIME math exams once it began retaining algebraic insights across questions. Similarly, GPT-4o’s success rate on the Game of 24 puzzle increased from about 10% to 99% after the model discovered and reused a Python-based solution. In tasks prone to arithmetic mistakes, such as balancing equations, DC enabled GPT-4o and Claude to reach near-perfect accuracy by recalling previously validated code, whereas their baselines stagnated around 50% . Beyond arithmetic challenges, DC yields notable accuracy gains on knowledge-demanding tasks. Claude achieved a 9% improvement in GPQA-Diamond and an 8% boost on MMLU-Pro Engineering and Physics problems. Crucially, DC’s memory is self-curated, focusing on concise, transferable snippets rather than entire transcripts, thereby facilitating meta-learning and avoiding context ballooning. Unlike fine-tuning or static retrieval methods, DC adapts LMs’ problem-solving skills on the fly, without modifying their underlying parameters, and offers a practical approach for continuously refining responses and cutting routine errors. Overall, our findings present DC as a promising approach for augmenting LMs with persistent memory, bridging the divide between isolated inference events and the cumulative, experience-driven learning characteristic of human cognition.*  \nFigure 1: Comparison of different baselines and Dynamic Cheatsheet (DC) variants on challenging reasoning benchmarks, including AIME exams and GPQA-Diamond. Baseline represents a standard prompting approach with minimal guidance, while DC-∅ (a stronger baseline) contains explicit structured instructions for problem solving, as well as for Python code generation and execution, but lacks a memory component. Our proposed DC-Cu and DC-RS variants incorporate an evolving, text-based memory to enhance inference-time learning. Results (accuracy,%) demonstrate substantial improvements, with Claude 3.5 Sonnet gaining 27% on AIME 2024 and 30% on AIME 2025 under DC-Cu. In Game of 24, GPT-4o leaps from 10%(baseline) to 99% under DC-RS, reflecting its ability to retain and apply Python-based solutions efficiently. Similarly, Claude 3.5 Sonnet’s accuracy more than doubles in Math Equation Solver, reaching 98% . Overall, these findings highlight the impact of test-time learning through controlled memory augmentation and efficient retrieval.  \n1 Stanford University 2Together AI. 􀀀 Correspondence to: msuzgun@stanford.edu and jamesz@stanford.edu.  \n* § We release all our data, results, and code at [http://github.com/suzgunmirac/dynamic-cheatsheet](http://github.com/suzgunmirac/dynamic-cheatsheet).  \n1. Introduction  \nModern large language models (LLMs) can tackle complex reasoning tasks, answer various questions, and generate extensive texts. Yet they still suffer from one critical limitation: once deployed, these models are fixed prior to deployment and typically retain no explicit or implicit memory of past questions, successes, or mistakes during inference. They approach each new problem de novo, often r","cbCailf4ZJBQCHWl","https://ap.wps.com/l/cbCailf4ZJBQCHWl","pdf",3794457,22,"English","# Introduction\n## Dynamic Cheatsheet overview\n## Workflow and memory curation\n## Evaluation across benchmarks","[{\"question\":\"What problem does Dynamic Cheatsheet address for language models?\",\"answer\":\"It addresses the limitation that deployed language models typically have no persistent memory of prior questions, successes, or mistakes during inference, causing repeated re-derivation of the same insights and repeated errors.\"},{\"question\":\"How does DC work during inference time?\",\"answer\":\"DC first consults an external memory to retrieve relevant stored lessons, generates a solution by combining retrieved insights with internal reasoning, and then updates the memory via a curation step based on whether the approach is correct or an error occurs.\"},{\"question\":\"Does DC require fine-tuning or ground-truth labels?\",\"answer\":\"No. DC is designed to improve models at inference time without gradient-based parameter updates and without needing explicit ground-truth labels or human feedback.\"}]","Dynamic Cheatsheet - Dynamic Cheatsheet - Test-Time Learning with Adaptive Memory | PDF",1791928580,8]