[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-160431-en":3,"doc-seo-160431-105":31,"detail-sidebar-cat-0-en-105":92},{"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":20,"is_deleted":4,"is_public":21,"is_downloadable":21,"audit_status":21,"page_count":22,"language":23,"language_code":24,"site_id":25,"html_lang":24,"table_of_contents":26,"faqs":27,"seo_title":28,"seo_description":14,"update_tm":29,"read_time":30},160431,687207024478,"Mia  ","https://ap-avatar.wpscdn.com/davatar_a8503ba1806abce46bf441b54a3ca4cd",8,"Research & Report","Learning to Plan by Updating Natural Language","Large Language Models (LLMs) often perform well on basic language tasks, yet complex tasks require a structured plan that guides step-by-step solution generation. Existing self-generated plans may be factually wrong or incomplete, so high-quality plans must include correct step sequences and behavioral instructions to avoid mistakes. Learning to Plan introduces two phases: iteratively updating task plans from training error feedback, then using the learned plan to guide inference on test sets.","Learning to Plan by Updating Natural Language  \nYiduo Guo 1 ,, Yaobo Liang2 , Chenfei Wu2 , Wenshan Wu2 , Dongyan Zhao 1 , Nan Duan2  \n1Wangxuan Institute of Computer Technology, Peking University  \n2Microsoft Research Asia  \n[yiduo@stu.pku.edu.cn](yiduo@stu.pku.edu.cn) , [zhaodongyan@pku.edu.cn](zhaodongyan@pku.edu.cn),{yaobo.liang, chenfei.wu, wenshan.wu, [nanduan}@microsoft.com](nanduan}@microsoft.com)  \narXiv :2304 . 10464v4 [ cs .CL] 13 Dec 2023  \nAbstract  \nLarge Language Models (LLMs) have shown remarkable performance in various basic natural language tasks. For completing the complex task, we still need a plan for the task to guide LLMs to generate the specific solutions step by step. LLMs can directly generate task plans, but these plans may still contain factual errors or are incomplete. A high-quality task plan contains correct step-by-step solutions for solving all situations and behavioral instructions for avoiding mistakes. To obtain it, we propose the Learning to Plan method, which involves two phases: (1) In the first learning task plan phase, it iteratively updates the task plan with new step-by-step solutions and behavioral instructions, which are obtained by prompting LLMs to derive from training error feedback.  \n(2) In the subsequent test phase, the LLM uses the learned task plan to guide the inference of LLM on the test set. We demonstrate the effectiveness of our method on the five different reasoning type tasks (8 datasets) . Further, our analysis experiment shows that the task plan learned by one LLM can directly guide another LLM to improve its performance, which reveals a new transfer learning paradigm.1  \n1 Introduction  \nLarge Language Models (LLMs), such as ChatGPT and GPT-4 (OpenAI, 2023), have recently achieved strong zero-shot/few-shot performance on various natural language tasks, such as generating passages (Bang et al., 2023), generating code (Liu et al., 2023a), and solving grade school math problems (Qin et al., 2023) . LLMs can further learn new basic abilities by connecting them with millions of APIs like TaskMatrix.AI (Lianget al., 2023 ; Wu et al., 2023) or new tools like ToolFormer (Schick et al., 2023) . However, LLMs still struggle to complete complex tasks, such as writing a long novel (Yang et al., 2022), coding for a  \n1We release the code at [https://github.com/](https://github.com/)[ ](https://github.com/)Eureka6174/LearnNLPlan  \n\n| Task instruction (prompt): Calculating the sin/cos value of an angle based on the length of legs and other conditions in a triangle. |\n| --- |\n| Input: In a triangle ABC, AC is 8 units, and BC is 15 units, what is cos\u003CABC? |\n| Label: ~~ 1~~1~~5~~7 |\n| Zero-shot chain of thought: Let’s think step by step. Corresponding solution of the input: Apply the cosine\u003Cbr>rule: cos( ̸ ABC) = ~~8~~2~~ ~~~~15~~82−15~~15~~2 = ~~6~~24~~4~~0\u003Cbr>Simplify: cos( ̸ ABC) = ~~4~~15 |\n| Task plan learned by our method: There are two solutions for finding the value of an angle in a triangle. Solution 1 involves using the Law of Cosines to calculate cos( ̸ ABC) .\u003Cbr>Solution 2 involves using the Pythagorean Theorem to find the length of the third side of the triangle and then using the definition of sine to calculate sin( ̸ ABC) . To find the length of the hypotenuse of a right triangle using the Pythagorean Theorem, use the formula (AC)2 +(BC)2 =(AB)2 . python functions can also be used to calculate these values.\u003Cbr>Corresponding solution of the input: Using the Pythagorean theorem, we can have AB = √82 + 152 = 17. Then, based on the definition of cosines, we can have cos \u003C ABC = ~~1~~1~~5~~7 |\n\nTable 1: We list the typical elements of a task.  \nlarge project (Orlando), and solving complex math problems (Frieder et al., 2023) . This indicates that knowing every basic step/capability is insufficient to complete complex tasks-we still require a task plan consisting of step-by-step solutions and behavioral instructions to solve the complex task. We can employ human experts to ","cbCaibiQALqmtt9d","https://ap.wps.com/l/cbCaibiQALqmtt9d","pdf",911660,3,1,36,"English","en",105,"# Abstract\n## Introduction\n## Learning to Plan Method\n## Experiments and Results","[{\"question\":\"Why do LLMs need task plans for complex problems?\",\"answer\":\"For complex tasks, knowing basic abilities is insufficient. A task plan provides step-by-step solutions and behavioral instructions that guide the model toward correct outputs while reducing mistakes.\"},{\"question\":\"How does the Learning to Plan method update the task plan?\",\"answer\":\"It collects wrong samples that the current plan cannot solve, then prompts the LLM to generate new step-by-step solutions. The best-performing update is selected on a validation set and appended to the existing plan, repeating until improvement stalls.\"},{\"question\":\"How is the learned task plan used during testing?\",\"answer\":\"During the test phase, the learned task plan is inserted into the prompt to guide the LLM’s inference on the test set, improving performance across multiple reasoning datasets.\"}]","Learning to Plan by Updating Natural Language | PDF",1788063343,91,{"code":4,"msg":32,"data":33},"ok",{"site_id":25,"language":24,"slug":34,"title":13,"keywords":35,"description":14,"schema_data":36,"social_meta":87,"head_meta":89,"extra_data":91,"updated_unix":29},"learning-to-plan-by-updating-natural-language","",{"@graph":37,"@context":86},[38,54,69],{"@type":39,"itemListElement":40},"BreadcrumbList",[41,45,49,51],{"item":42,"name":43,"@type":44,"position":21},"https://docshare.wps.com","Home","ListItem",{"item":46,"name":47,"@type":44,"position":48},"https://docshare.wps.com/document/","Document",2,{"item":50,"name":12,"@type":44,"position":20},"https://docshare.wps.com/document/research-report/",{"item":52,"name":13,"@type":44,"position":53},"https://docshare.wps.com/document/learning-to-plan-by-updating-natural-language/160431/",4,{"url":52,"name":13,"@type":55,"author":56,"headline":13,"publisher":58,"fileFormat":61,"inLanguage":24,"description":14,"dateModified":62,"datePublished":63,"encodingFormat":61,"isAccessibleForFree":64,"interactionStatistic":65},"DigitalDocument",{"name":9,"@type":57},"Person",{"url":42,"name":59,"@type":60},"DocShare","Organization","application/pdf","2026-09-06","2026-08-30",true,{"@type":66,"interactionType":67,"userInteractionCount":20},"InteractionCounter",{"@type":68},"ViewAction",{"@type":70,"mainEntity":71},"FAQPage",[72,78,82],{"name":73,"@type":74,"acceptedAnswer":75},"Why do LLMs need task plans for complex problems?","Question",{"text":76,"@type":77},"For complex tasks, knowing basic abilities is insufficient. A task plan provides step-by-step solutions and behavioral instructions that guide the model toward correct outputs while reducing mistakes.","Answer",{"name":79,"@type":74,"acceptedAnswer":80},"How does the Learning to Plan method update the task plan?",{"text":81,"@type":77},"It collects wrong samples that the current plan cannot solve, then prompts the LLM to generate new step-by-step solutions. The best-performing update is selected on a validation set and appended to the existing plan, repeating until improvement stalls.",{"name":83,"@type":74,"acceptedAnswer":84},"How is the learned task plan used during testing?",{"text":85,"@type":77},"During the test phase, the learned task plan is inserted into the prompt to guide the LLM’s inference on the test set, improving performance across multiple reasoning datasets.","https://schema.org",{"og:url":52,"og:type":88,"og:title":13,"og:site_name":59,"og:description":14},"article",{"robots":90,"canonical":52},"index,follow",{"doc_id":7,"site_id":25},{"code":4,"msg":5,"data":93},[94,98,102,106,111,116,121,124,129,132,136],{"id":21,"doc_module":4,"doc_module_name":47,"category_name":95,"show_sort_weight":96,"slug":97},"Story & Novel",90,"story-novel",{"id":48,"doc_module":4,"doc_module_name":47,"category_name":99,"show_sort_weight":100,"slug":101},"Literature",80,"literature",{"id":53,"doc_module":4,"doc_module_name":47,"category_name":103,"show_sort_weight":104,"slug":105},"Exam",70,"exam",{"id":107,"doc_module":4,"doc_module_name":47,"category_name":108,"show_sort_weight":109,"slug":110},5,"Comic",60,"comic",{"id":112,"doc_module":4,"doc_module_name":47,"category_name":113,"show_sort_weight":114,"slug":115},6,"Technology",50,"technology",{"id":117,"doc_module":4,"doc_module_name":47,"category_name":118,"show_sort_weight":119,"slug":120},7,"Healthcare",40,"healthcare",{"id":11,"doc_module":4,"doc_module_name":47,"category_name":12,"show_sort_weight":122,"slug":123},30,"research-report",{"id":125,"doc_module":4,"doc_module_name":47,"category_name":126,"show_sort_weight":127,"slug":128},9,"Religion & Spirituality",20,"religion-spirituality",{"id":127,"doc_module":4,"doc_module_name":47,"category_name":130,"show_sort_weight":127,"slug":131},"World Cup","world-cup",{"id":133,"doc_module":4,"doc_module_name":47,"category_name":134,"show_sort_weight":133,"slug":135},10,"Lifestyle","lifestyle",{"id":137,"doc_module":4,"doc_module_name":47,"category_name":138,"show_sort_weight":107,"slug":139},19,"General","general"]