[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-seo-327492-105":3,"detail-sidebar-cat-0-en-105":81,"doc-detail-327492-en":131},{"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":74,"head_meta":76,"extra_data":78,"updated_unix":80},105,"en","evopolicygym-evaluating-autonomous-policy-evolution-in-interactive-environments","EvoPolicyGym - Evaluating Autonomous Policy Evolution in Interactive Environments","","Autonomous agents are increasingly expected to improve executable policies through feedback, yet many evaluations reduce this to final scores or blend it with software-engineering progress. The document introduces Autonomous Policy Evolution, a controlled harness–model setting where an agent repeatedly edits an executable policy system within a fixed interaction budget. EvoPolicyGym instantiates the setting using compact interactive RL environments to assess iterative policy improvement. On the suite, GPT-5.5 achieves the strongest aggregate rank and top-two results across all 16 environments, and provides trajectory-level diagnostics for budget allocation and feedback-to-parameter conversion, highlighting the mechanisms needed under bounded feedback.",{"@graph":14,"@context":73},[15,34,56],{"@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/document/","Document",2,{"item":28,"name":29,"@type":21,"position":30},"https://docshare.wps.com/document/research-report/","Research & Report",3,{"item":32,"name":10,"@type":21,"position":33},"https://docshare.wps.com/document/evopolicygym-evaluating-autonomous-policy-evolution-in-interactive-environments/327492/",4,{"url":32,"name":10,"@type":35,"image":36,"author":41,"headline":10,"publisher":44,"fileFormat":47,"inLanguage":8,"description":12,"dateModified":48,"datePublished":49,"encodingFormat":47,"isAccessibleForFree":50,"interactionStatistic":51},"DigitalDocument",{"url":37,"@type":38,"width":39,"height":40},"https://docshare.wps.com/thumbnails/evopolicygym-evaluating-autonomous-policy-evolution-in-interactive-environments/327492.png","ImageObject",300,407,{"name":42,"@type":43},"Hazel","Person",{"url":19,"name":45,"@type":46},"DocShare","Organization","application/pdf","2026-11-01","2026-09-21",true,{"@type":52,"interactionType":53,"userInteractionCount":55},"InteractionCounter",{"@type":54},"ViewAction",12,{"@type":57,"mainEntity":58},"FAQPage",[59,65,69],{"name":60,"@type":61,"acceptedAnswer":62},"What evaluation gap does the document address?","Question",{"text":63,"@type":64},"It highlights that existing evaluations often collapse autonomous improvement into a single final score or confuse it with open-ended software-engineering progress, masking both outcome and process-level failures.","Answer",{"name":66,"@type":61,"acceptedAnswer":67},"How does Autonomous Policy Evolution define the evaluation setting?",{"text":68,"@type":64},"An agent repeatedly revises an executable decision policy using feedback from prior deployments, with the goal measured by held-out return on hidden validation under a fixed interaction budget.",{"name":70,"@type":61,"acceptedAnswer":71},"What key results are reported on EvoPolicyGym?",{"text":72,"@type":64},"GPT-5.5 obtains the highest aggregate rank score and top-two performance on all 16 environments, while other agents achieve more limited cross-environment coverage.","https://schema.org",{"og:url":32,"og:type":75,"og:title":10,"og:site_name":45,"og:description":12},"article",{"robots":77,"canonical":32},"index,follow",{"doc_id":79,"site_id":7},327492,1790465340,{"code":4,"msg":82,"data":83},"success",[84,88,92,96,101,106,111,115,120,123,127],{"id":22,"doc_module":4,"doc_module_name":25,"category_name":85,"show_sort_weight":86,"slug":87},"Story & Novel",90,"story-novel",{"id":26,"doc_module":4,"doc_module_name":25,"category_name":89,"show_sort_weight":90,"slug":91},"Literature",80,"literature",{"id":33,"doc_module":4,"doc_module_name":25,"category_name":93,"show_sort_weight":94,"slug":95},"Exam",70,"exam",{"id":97,"doc_module":4,"doc_module_name":25,"category_name":98,"show_sort_weight":99,"slug":100},5,"Comic",60,"comic",{"id":102,"doc_module":4,"doc_module_name":25,"category_name":103,"show_sort_weight":104,"slug":105},6,"Technology",50,"technology",{"id":107,"doc_module":4,"doc_module_name":25,"category_name":108,"show_sort_weight":109,"slug":110},7,"Healthcare",40,"healthcare",{"id":112,"doc_module":4,"doc_module_name":25,"category_name":29,"show_sort_weight":113,"slug":114},8,30,"research-report",{"id":116,"doc_module":4,"doc_module_name":25,"category_name":117,"show_sort_weight":118,"slug":119},9,"Religion & Spirituality",20,"religion-spirituality",{"id":118,"doc_module":4,"doc_module_name":25,"category_name":121,"show_sort_weight":118,"slug":122},"World Cup","world-cup",{"id":124,"doc_module":4,"doc_module_name":25,"category_name":125,"show_sort_weight":124,"slug":126},10,"Lifestyle","lifestyle",{"id":128,"doc_module":4,"doc_module_name":25,"category_name":129,"show_sort_weight":97,"slug":130},19,"General","general",{"code":4,"msg":82,"data":132},{"doc_id":79,"user_id":133,"nickname":42,"user_avatar":134,"doc_module":4,"category_id":112,"category_name":29,"doc_title":10,"doc_description":12,"doc_content":135,"file_id":136,"file_url":137,"file_type":138,"file_size":139,"view_count":55,"is_deleted":4,"is_public":22,"is_downloadable":22,"audit_status":22,"page_count":140,"language":141,"language_code":8,"site_id":7,"html_lang":8,"table_of_contents":142,"faqs":143,"seo_title":144,"seo_description":12,"update_tm":145,"read_time":99,"preview_page_count":4},137441390410,"https://ap-avatar.wpscdn.com/avatar/2000252f4ab5702993?_k=1776741390130283984","arXiv :2607 .02440v 1 [ cs .AI] 2 Jul 2026  \nEvoPolicyGym: Evaluating Autonomous Policy Evolution in Interactive Environments  \nZhilin Wang1,* , Han Song2,* , Runzhe Zhan3,* Jusen Du4 , Jiacheng Chen2 , Tianle Li2 , Qingyu Yin5 , Yulun Wu5 Zhennan Shen9 , Tong Zhu6 , Yanshu Li7 , Guanjie Chen9 Derek F. Wong3 , Yafu Li2,†, Yu Cheng2,†, Yang Yang9,†  \n1University of Science and Technology of China 2The Chinese University of Hong Kong 3University of Macau 4Tsinghua University 5 Zhejiang University 6Soochow University 7Brown University  \n9Shanghai Jiao Tong University  \n* Equal contribution †Corresponding authors  \nAbstract. Autonomous agents are increasingly expected to improve executable policies through feedback, yet existing evaluations often collapse this process into a final score or confound it with open-ended software-engineering progress. We introduce Autonomous Policy Evolution, a controlled evaluation setting in which a harness–model agent repeatedly edits an executable policy system under a fixed interaction budget. We instantiate this setting in EvoPolicyGym, a benchmark built from compact interactive RL environments that evaluateshow agents iteratively improve explored policies. On the EvoPolicyGym suite, GPT- 5.5 achieves the strongest aggregate rank score and top-two performance on all 16 environments. Beyond leaderboard results, EvoPolicyGym also provides trajectory-level diagnostics that distinguish how agents allocate budget, convert feedback into parametric tuning. These analyses show that strong autonomous policy evolution depends not only on isolated task wins, but on discovering task-appropriate mechanismsand refining policies under bounded feedback.  \n§ EvoPolicyGym õ EvoPolicyGym-Exp-data 􀂀 EvoPolicyGym . io  \n1 Introduction  \nAutonomous agents are expected to improve through feedback rather than produce a single fixed output. Modern coding agents are able to call tools, observe failures, and revise executable artifacts over long horizons (Yang et al., 2024; Wang et al., 2024b;a), while self-improvement systems show that language models can use reflection to refine answers across attempts (Shinn et al., 2023; Madaan et al., 2023; Novikov et al., 2025) . However, this broader capability is hard to evaluate because improvement is both an outcome anda process: final scores can hide blind retries, overfitting to visible feedback, brittle special cases, missing verification, and other trajectory-level failure modes (Lu et al., 2024; Majgaonkar et al., 2025; Tang et al., 2026) . Fully open-ended engineering tasks add further confounders, including evolving specifications and  \nsoftware-maintenance quality (Chen et al., 2026; Xu et al., 2026; Hamblin et al., 2026; Orlanski et al., 2026) . We therefore need a controlled setting that isolates an agent’s ability to convert bounded environment feedback into generalizable improvements of an executable policy, while retaining the iterative decisions that make autonomous improvement difficult.  \nWe address this gap by formalizing Autonomous Policy Evolution, a problem in which an agent repeatedly revises an executable decision policy using feedback from prior deployments. Formally, the observable object is the sequence of submitted policy systems and train-feedback records, whereas the outcome is the held-out return of the checkpoint selected on hidden validation. The goal is not only to maximize observed performance alone, but to produce a policy that generalizes to environment instances after limited interaction. The bounded budget is therefore part of the capability being measured: it requires each agent to choose what information to acquire, when to explore or exploit, and how efficiently to convert sparse behavioral evidence into robust policy improvement.  \nTo evaluate this problem, we instantiate Autonomous Policy Evolution in EvoPolicyGym, a controlled benchmark built from compact interactive environments. Unlike open-ended engineering benchmarks (Chi et al., 20","cbCailx24qUir5aa","https://ap.wps.com/l/cbCailx24qUir5aa","pdf",1395080,24,"English","# 1 Introduction\n## Autonomous Policy Evolution and evaluation gap\n## EvoPolicyGym benchmark protocol\n## Preliminary experiments and results\n## Contributions and diagnostic analysis","[{\"question\":\"What evaluation gap does the document address?\",\"answer\":\"It highlights that existing evaluations often collapse autonomous improvement into a single final score or confuse it with open-ended software-engineering progress, masking both outcome and process-level failures.\"},{\"question\":\"How does Autonomous Policy Evolution define the evaluation setting?\",\"answer\":\"An agent repeatedly revises an executable decision policy using feedback from prior deployments, with the goal measured by held-out return on hidden validation under a fixed interaction budget.\"},{\"question\":\"What key results are reported on EvoPolicyGym?\",\"answer\":\"GPT-5.5 obtains the highest aggregate rank score and top-two performance on all 16 environments, while other agents achieve more limited cross-environment coverage.\"}]","EvoPolicyGym - Evaluating Autonomous Policy Evolution in Interactive Environments | PDF",1789977324]