Hugging Face Daily Papers · June 16, 2026 · 4 min read

FastContext: Training Efficient Repository Explorer for Coding Agents

Mirrored from Hugging Face Daily Papers for archival readability. Support the source by reading on the original site.

Like Read original ↗

Can an explore subagent rise the performance of main agent?\n","updatedAt":"2026-06-16T02:19:56.175Z","author":{"_id":"65db54e5ab2f64915c0b9cf0","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/65db54e5ab2f64915c0b9cf0/fMnaW3h3gFx84TAa3H3sH.png","fullname":"Shaoqiu Zhang","name":"qiushao","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":3,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.632767379283905},"editors":["qiushao"],"editorAvatarUrls":["https://cdn-avatars.huggingface.co/v1/production/uploads/65db54e5ab2f64915c0b9cf0/fMnaW3h3gFx84TAa3H3sH.png"],"reactions":[{"reaction":"❤️","users":["maoquan-ms","YerbaPage","qiushao"],"count":3}],"isReport":false}},{"id":"6a30c7908e258cbed42169c8","author":{"_id":"645b0c3ec35da9c7afd95421","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/645b0c3ec35da9c7afd95421/vYBrCDagHsXAo6J2p-uG0.jpeg","fullname":"Yuling","name":"YerbaPage","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":107,"isUserFollowing":false},"createdAt":"2026-06-16T03:48:32.000Z","type":"comment","data":{"edited":false,"hidden":false,"latest":{"raw":"We develop FastContext to tackle context pollution by offloading exploration. \n\nFor the bigger picture of agent context compression, feel free to check our latest survey!\n\n📄 https://doi.org/10.20944/preprints202605.2065.v1\n⭐ https://github.com/YerbaPage/Awesome-Agent-Context-Compression","html":"We develop FastContext to tackle context pollution by offloading exploration. \nFor the bigger picture of agent context compression, feel free to check our latest survey!\n📄 <a href=\"https://doi.org/10.20944/preprints202605.2065.v1\" rel=\"nofollow\">https://doi.org/10.20944/preprints202605.2065.v1</a> ⭐ <a href=\"https://github.com/YerbaPage/Awesome-Agent-Context-Compression\" rel=\"nofollow\">https://github.com/YerbaPage/Awesome-Agent-Context-Compression</a>\n","updatedAt":"2026-06-16T03:48:32.042Z","author":{"_id":"645b0c3ec35da9c7afd95421","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/645b0c3ec35da9c7afd95421/vYBrCDagHsXAo6J2p-uG0.jpeg","fullname":"Yuling","name":"YerbaPage","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":107,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.6977500915527344},"editors":["YerbaPage"],"editorAvatarUrls":["https://cdn-avatars.huggingface.co/v1/production/uploads/645b0c3ec35da9c7afd95421/vYBrCDagHsXAo6J2p-uG0.jpeg"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2606.14066","authors":[{"_id":"6a2fb3f3a0d4daae4285f977","user":{"_id":"65db54e5ab2f64915c0b9cf0","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/65db54e5ab2f64915c0b9cf0/fMnaW3h3gFx84TAa3H3sH.png","isPro":false,"fullname":"Shaoqiu Zhang","user":"qiushao","type":"user","name":"qiushao"},"name":"Shaoqiu Zhang","status":"claimed_verified","statusLastChangedAt":"2026-06-16T12:07:54.567Z","hidden":false},{"_id":"6a2fb3f3a0d4daae4285f978","name":"Maoquan Wang","hidden":false},{"_id":"6a2fb3f3a0d4daae4285f979","user":{"_id":"645b0c3ec35da9c7afd95421","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/645b0c3ec35da9c7afd95421/vYBrCDagHsXAo6J2p-uG0.jpeg","isPro":false,"fullname":"Yuling","user":"YerbaPage","type":"user","name":"YerbaPage"},"name":"Yuling Shi","status":"claimed_verified","statusLastChangedAt":"2026-06-15T12:18:09.430Z","hidden":false},{"_id":"6a2fb3f3a0d4daae4285f97a","name":"Yuhang Wang","hidden":false},{"_id":"6a2fb3f3a0d4daae4285f97b","name":"Xiaodong Gu","hidden":false},{"_id":"6a2fb3f3a0d4daae4285f97c","name":"Yongqiang Yao","hidden":false},{"_id":"6a2fb3f3a0d4daae4285f97d","name":"Rao Fu","hidden":false},{"_id":"6a2fb3f3a0d4daae4285f97e","name":"Shengyu Fu","hidden":false}],"publishedAt":"2026-06-12T00:00:00.000Z","submittedOnDailyAt":"2026-06-16T00:00:00.000Z","title":"FastContext: Training Efficient Repository Explorer for Coding Agents","submittedOnDailyBy":{"_id":"65db54e5ab2f64915c0b9cf0","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/65db54e5ab2f64915c0b9cf0/fMnaW3h3gFx84TAa3H3sH.png","isPro":false,"fullname":"Shaoqiu Zhang","user":"qiushao","type":"user","name":"qiushao"},"summary":"Large Language Model (LLM) coding agents have achieved strong results on software engineering tasks, yet repository exploration remains a major bottleneck: locating relevant code consumes substantial token budget and pollutes the agent's context with irrelevant snippets. In most agents, the same model explores the repository and solves the task, leaving exploratory reads and searches in the solver's history. We present FastContext, a dedicated exploration subagent that separates repository exploration from solving. Invoked on demand, FastContext issues parallel tool calls and returns concise file paths and line ranges as focused context. FastContext is powered by specialized exploration models spanning 4B--30B parameters. We bootstrap them from strong reference-model trajectories and refine them with task-grounded rewards for broad first-turn search, multi-turn evidence gathering, and precise citation generation. Across SWE-bench Multilingual, SWE-bench Pro, and SWE-QA, integrating FastContext into Mini-SWE-Agent improves end-to-end resolution rates up to 5.5\\% while reducing coding-agent token consumption up to 60\\%, with marginal overhead. These results show that repository exploration can be separated from solving and handled effectively by specialized models. Code and data: https://github.com/microsoft/fastcontext","upvotes":47,"discussionId":"6a2fb3f4a0d4daae4285f97f","projectPage":"https://huggingface.co/microsoft/FastContext-1.0-4B-SFT","githubRepo":"https://github.com/microsoft/fastcontext","githubRepoAddedBy":"user","ai_summary":"FastContext separates repository exploration from code solving in LLM agents using specialized exploration models that reduce token consumption and improve resolution rates.","ai_keywords":["large language model","coding agents","repository exploration","context management","exploration subagent","tool calls","specialized exploration models","reference-model trajectories","task-grounded rewards","first-turn search","multi-turn evidence gathering","precise citation generation","SWE-bench Multilingual","SWE-bench Pro","SWE-QA","Mini-SWE-Agent"],"ai_summary_model":"Qwen/Qwen2.5-Coder-32B-Instruct","githubStars":312,"organization":{"_id":"5e6485f787403103f9f1055e","name":"microsoft","fullname":"Microsoft","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/1583646260758-5e64858c87403103f9f1055d.png"}},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"645b0c3ec35da9c7afd95421","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/645b0c3ec35da9c7afd95421/vYBrCDagHsXAo6J2p-uG0.jpeg","isPro":false,"fullname":"Yuling","user":"YerbaPage","type":"user"},{"_id":"68b9405bc6bdb23cfbd73a8f","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/no-auth/A9wQmyLIC7vkAkjsaOHFz.png","isPro":false,"fullname":"Haowen Gong","user":"Foreverdream","type":"user"},{"_id":"6a2779eae75f45d33e88eef4","avatarUrl":"/avatars/5ab8738921a135d561793996a2c4b249.svg","isPro":false,"fullname":"Xy Gong","user":"kanata-0527","type":"user"},{"_id":"68ff4946425e4f340838fa31","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/68ff4946425e4f340838fa31/kGN_-9vsVuVTxStyQ108e.jpeg","isPro":false,"fullname":"Qin Haoxiang","user":"lithium-flower","type":"user"},{"_id":"69bfa7b62a6ad8b64110b6f6","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/69bfa7b62a6ad8b64110b6f6/2JTjI0iRlxeUN2uEmrCnx.jpeg","isPro":false,"fullname":"Ting","user":"QXQ666666","type":"user"},{"_id":"67640f123ebd56f692b00d9f","avatarUrl":"/avatars/db00af73849efa7a48639503389e1b96.svg","isPro":false,"fullname":"jack","user":"kimi000","type":"user"},{"_id":"6579e083bd9ea8b29c4901cb","avatarUrl":"/avatars/d60e0c32340ce2adc7ac84eeb63669a0.svg","isPro":false,"fullname":"xxxxxz","user":"xxx405","type":"user"},{"_id":"66a85c939836830cc1d3300c","avatarUrl":"/avatars/2a7c638644f15ce091fb688ec0da781f.svg","isPro":false,"fullname":"Tianao Zhang","user":"ZTAZTAZTA","type":"user"},{"_id":"6a27e00fbef79e443ba5aac0","avatarUrl":"/avatars/3a802e6d8d9298c782206750d4fa91b5.svg","isPro":false,"fullname":"Long Zhaofei","user":"zhaofeiisfading","type":"user"},{"_id":"6977604545db3773ec66d7ab","avatarUrl":"/avatars/1ae385499735d389f0505051d781ee54.svg","isPro":false,"fullname":"Cheng Yichun","user":"Oscillater","type":"user"},{"_id":"6a277c67eacfefb2ac5856b6","avatarUrl":"/avatars/2f74e0ac92a12dd81243479b6edc7efe.svg","isPro":false,"fullname":"d","user":"djenqbd","type":"user"},{"_id":"6a2a49426e44efe7028b7e6e","avatarUrl":"/avatars/c9bac6a8f9d3c004dce0cc504009bc13.svg","isPro":false,"fullname":"Maoquan","user":"maoquan-ms","type":"user"}],"acceptLanguages":["en"],"dailyPaperRank":0,"organization":{"_id":"5e6485f787403103f9f1055e","name":"microsoft","fullname":"Microsoft","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/1583646260758-5e64858c87403103f9f1055d.png"},"markdownContentUrl":"https://huggingface.co/buckets/huggingchat/papers-content/resolve/2606/2606.14066.md","query":{}}">

Papers

arxiv:2606.14066

FastContext: Training Efficient Repository Explorer for Coding Agents

Published on Jun 12

· Submitted by

Shaoqiu Zhang on Jun 16

Microsoft

Upvote

Authors:

Shaoqiu Zhang ,

Yuling Shi ,

Abstract

FastContext separates repository exploration from code solving in LLM agents using specialized exploration models that reduce token consumption and improve resolution rates.

Generated by Qwen/Qwen2.5-Coder-32B-Instruct

Large Language Model (LLM) coding agents have achieved strong results on software engineering tasks, yet repository exploration remains a major bottleneck: locating relevant code consumes substantial token budget and pollutes the agent's context with irrelevant snippets. In most agents, the same model explores the repository and solves the task, leaving exploratory reads and searches in the solver's history. We present FastContext, a dedicated exploration subagent that separates repository exploration from solving. Invoked on demand, FastContext issues parallel tool calls and returns concise file paths and line ranges as focused context. FastContext is powered by specialized exploration models spanning 4B--30B parameters. We bootstrap them from strong reference-model trajectories and refine them with task-grounded rewards for broad first-turn search, multi-turn evidence gathering, and precise citation generation. Across SWE-bench Multilingual, SWE-bench Pro, and SWE-QA, integrating FastContext into Mini-SWE-Agent improves end-to-end resolution rates up to 5.5\% while reducing coding-agent token consumption up to 60\%, with marginal overhead. These results show that repository exploration can be separated from solving and handled effectively by specialized models. Code and data: https://github.com/microsoft/fastcontext