[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-83966-en":3,"doc-seo-83966-105":30,"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":13,"seo_description":14,"update_tm":28,"read_time":29},83966,687197207639,"Asher","https://ap-avatar.wpscdn.com/davatar_a8503ba1806abce46bf441b54a3ca4cd",8,"Research & Report","KAT-Coder-V2.5 Technical Report","KAT-Coder-V2.5 is a coding-focused agentic model designed to act autonomously inside real, executable repositories rather than functioning as a single-turn code generator. Performance is limited less by model size than by scarce reproducible environments, verifiable rewards, and high-value training trajectories. AutoBuilder reconstructs multilingual repositories into sandboxed, fail-to-pass and pass-to-pass verified environments, regenerating structured tasks and recovering near-miss behavior. KwaiClawEnv synthesizes large-scale tool-use trajectories from executable services and real task seeds, enabling stronger agentic tool use across SWE and agentic benchmarks.","arXiv :2607 .0547 1v 1 [ cs . SE] 6 Jul 2026  \nKAT-Coder-V2.5 Technical Report  \nKwaiKAT Team  \nAbstract  \nWe present KAT-Coder-V2.5, a coding-focused agentic model trained to act autonomously inside real, executable repositories rather than as a single-turn code generator. Its capability is bottlenecked less by model scale than by the scarcity of reproducible environments, verifiable rewards, and high-value trajectories, which we address with an end-to-end agentic post-training framework. AutoBuilder reconstructs multilingual repositories into sandboxed environments with fail-to-pass and pass-to-pass verification at scale, from which we regenerate self-contained task specifications, recover near-miss trajectories, and distill supervision through process-aware filtering, while KwaiClawEnv synthesizes large-scale tool-use trajectories from executable services and real task seeds. We further scale reinforcement learning with harness randomization, a reliability-hardened sandbox, an asymmetric actor–critic PPO with hindsightaugmented value estimation, and a harness-oriented reward framework, and unify SWE, Agent-Claw, and WebCoding experts via Multi-Teacher On-Policy Distillation. Across six software-engineering and agentic benchmarks, KAT-Coder-V2.5 delivers the best agentic tool-use result on PinchBench and ranks second only to the frontier Opus 4.8 on repository-level software engineering. Our service is available at [https://streamlake.com/product/kat-coder](https://streamlake.com/product/kat-coder).  \n\n| SWE-Bench Pro Coding\u003Cbr>\u003Cbr>  KAT-Coder-V2.5\u003Cbr>GLM-5.1 GLM-5.2\u003Cbr>Kimi K2.6 Opus 4.8 |  |\n| --- | --- |\n| KAT Claw Bench Claw |  |\n| \u003Cbr>KAT-Coder-V2.5\u003Cbr>GLM-5.1 GLM-5.2\u003Cbr>Kimi K2.6 Opus 4.8\u003Cbr> |  |\n\n\n| KAT Code Bench Coding\u003Cbr>\u003Cbr>KAT-Coder-V2.5\u003Cbr>GLM-5.1 GLM-5.2\u003Cbr>Kimi K2.6 Opus 4.8 |  |\n| --- | --- |\n\nTerminal-Bench 2.1 AA Coding Index  \nKAT-Coder-V2.5  \nGLM-5.1 GLM-5.2  \nKimi K2.6 Opus 4.8  \n\n| PinchBench Claw |  |\n| --- | --- |\n| \u003Cbr>KAT-Coder-V2.5\u003Cbr>GLM-5.2 Kimi K2.6\u003Cbr>Opus 4.8 |  |\n\n\n| SciCode AA Coding Index |  |\n| --- | --- |\n| \u003Cbr>KAT-Coder-V2.5\u003Cbr>GLM-5.1 GLM-5.2\u003Cbr>Kimi K2.6 Opus 4.8\u003Cbr> |  |\n\nFigure 1 | Results of KAT-Coder-V2.5 on various SWE and agent benchmarks.  \n1. Introduction  \nCoding models [1] have moved beyond passive code completion toward autonomous agents that understand requirements, navigate repositories, edit files, run tests, and iteratively repair their own failures [2, 3] . This shift redefines what capability means: a strong coding agent must combine repository understanding, tool use, long-horizon planning, engineering discipline, and verification-driven problem solving, not merely generate syntactically correct snippets. Requirements are stated in natural language, relevant logic is scattered across many files, and the agent must search, localize, design, edit, verify, and recover under limited context.  \nWe argue that the primary obstacle to building such agents is not model scale but the training infrastructure around it, which manifests as three concrete challenges.  \nScalable executable environments. Real repositories differ substantially in dependencies, build systems, testing frameworks, and runtime assumptions. Producing environments that are reproducible, executable, and truly verifiable is difficult at scale, and naive construction silently admits environments that never run the intended tests [4, 5] .  \nTrajectory quality beyond final rewards. Raw issues and pull-request descriptions are often incomplete, ambiguous, or inconsistent with the merged change. Worse, filtering trajectories by final test success alone is misleading: some passing trajectories rely on hard-coding, mechanism bypassing, or test-oriented shortcuts, while some failed trajectories still contain valuable search, localization, and repair behavior.  \nStable long-horizon RL. Long-horizon agentic reinforcement learning suffers from sparse rewards, unstable environment feedback, coarse credit assignment, ove","cbCaieZDnkKWftkQ","https://ap.wps.com/l/cbCaieZDnkKWftkQ","pdf",1325533,5,1,24,"English","en",105,"# Introduction\n## Training Challenges\n## Post-training Framework: AutoBuilder and KwaiClawEnv\n## Harness-Robust Reinforcement Learning","[{\"question\":\"What distinguishes KAT-Coder-V2.5 from a typical code generator?\",\"answer\":\"KAT-Coder-V2.5 is trained to operate autonomously inside real, executable repositories, navigating and verifying work rather than producing code in a single response.\"},{\"question\":\"What three training challenges does the report identify for building coding agents?\",\"answer\":\"The report highlights difficulty in scalable executable environments, problems with trajectory quality when judging only final rewards, and instability in stable long-horizon reinforcement learning.\"},{\"question\":\"How does the framework improve training signals and reliability?\",\"answer\":\"AutoBuilder builds verifiable sandbox environments with fail-to-pass and pass-to-pass checks, while KwaiClawEnv generates tool-use trajectories from executable services. Reinforcement learning further uses harness randomization, a reliability-hardened sandbox, and an asymmetric actor–critic PPO with harness-oriented rewards.\"}]",1784191710,60,{"code":4,"msg":31,"data":32},"ok",{"site_id":25,"language":24,"slug":33,"title":13,"keywords":34,"description":14,"schema_data":35,"social_meta":87,"head_meta":89,"extra_data":91,"updated_unix":28},"kat-coder-v25-technical-report","",{"@graph":36,"@context":86},[37,54,69],{"@type":38,"itemListElement":39},"BreadcrumbList",[40,44,48,51],{"item":41,"name":42,"@type":43,"position":21},"https://docshare.wps.com","Home","ListItem",{"item":45,"name":46,"@type":43,"position":47},"https://docshare.wps.com/document/","Document",2,{"item":49,"name":12,"@type":43,"position":50},"https://docshare.wps.com/document/research-report/",3,{"item":52,"name":13,"@type":43,"position":53},"https://docshare.wps.com/document/kat-coder-v25-technical-report/83966/",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":41,"name":59,"@type":60},"DocShare","Organization","application/pdf","2026-07-28","2026-07-16",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},"What distinguishes KAT-Coder-V2.5 from a typical code generator?","Question",{"text":76,"@type":77},"KAT-Coder-V2.5 is trained to operate autonomously inside real, executable repositories, navigating and verifying work rather than producing code in a single response.","Answer",{"name":79,"@type":74,"acceptedAnswer":80},"What three training challenges does the report identify for building coding agents?",{"text":81,"@type":77},"The report highlights difficulty in scalable executable environments, problems with trajectory quality when judging only final rewards, and instability in stable long-horizon reinforcement learning.",{"name":83,"@type":74,"acceptedAnswer":84},"How does the framework improve training signals and reliability?",{"text":85,"@type":77},"AutoBuilder builds verifiable sandbox environments with fail-to-pass and pass-to-pass checks, while KwaiClawEnv generates tool-use trajectories from executable services. Reinforcement learning further uses harness randomization, a reliability-hardened sandbox, and an asymmetric actor–critic PPO with harness-oriented rewards.","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},[94,98,102,106,109,114,119,122,127,130,134],{"id":21,"doc_module":4,"doc_module_name":46,"category_name":95,"show_sort_weight":96,"slug":97},"Story & Novel",90,"story-novel",{"id":47,"doc_module":4,"doc_module_name":46,"category_name":99,"show_sort_weight":100,"slug":101},"Literature",80,"literature",{"id":53,"doc_module":4,"doc_module_name":46,"category_name":103,"show_sort_weight":104,"slug":105},"Exam",70,"exam",{"id":20,"doc_module":4,"doc_module_name":46,"category_name":107,"show_sort_weight":29,"slug":108},"Comic","comic",{"id":110,"doc_module":4,"doc_module_name":46,"category_name":111,"show_sort_weight":112,"slug":113},6,"Technology",50,"technology",{"id":115,"doc_module":4,"doc_module_name":46,"category_name":116,"show_sort_weight":117,"slug":118},7,"Healthcare",40,"healthcare",{"id":11,"doc_module":4,"doc_module_name":46,"category_name":12,"show_sort_weight":120,"slug":121},30,"research-report",{"id":123,"doc_module":4,"doc_module_name":46,"category_name":124,"show_sort_weight":125,"slug":126},9,"Religion & Spirituality",20,"religion-spirituality",{"id":125,"doc_module":4,"doc_module_name":46,"category_name":128,"show_sort_weight":125,"slug":129},"World Cup","world-cup",{"id":131,"doc_module":4,"doc_module_name":46,"category_name":132,"show_sort_weight":131,"slug":133},10,"Lifestyle","lifestyle",{"id":135,"doc_module":4,"doc_module_name":46,"category_name":136,"show_sort_weight":20,"slug":137},19,"General","general"]