Hugging Face Daily Papers · August 27, 2026 · 4 min read

Agent-G^2: Gaussian Guidance for Agentic Reinforcement Learning

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

Like Read original ↗

We propose Agent-G^2, a Gaussian guidance framework that draws the depth per task from a Gaussian whose center and spread are estimated online from rollouts already collected for policy optimization, requiring no probe rollouts or learned depth predictor.</p>\n","updatedAt":"2026-08-27T02:18:31.741Z","author":{"_id":"676127cf11b19ea602bb202a","avatarUrl":"/avatars/dfd802a24bd63e509728159ebb1769f6.svg","fullname":"Zhengxi Lu","name":"LZXzju","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":13,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.830083429813385},"editors":["LZXzju"],"editorAvatarUrls":["/avatars/dfd802a24bd63e509728159ebb1769f6.svg"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2608.23318","authors":[{"_id":"6a8d0bb25add2537c32e9706","user":{"_id":"677d3523a6918748bf81d0e9","avatarUrl":"/avatars/c9961ac54d089efb36db20c421c2bea2.svg","isPro":false,"fullname":"wangzixuan","user":"wangzx1210","type":"user","name":"wangzx1210"},"name":"Zixuan Wang","status":"claimed_verified","statusLastChangedAt":"2026-08-25T08:13:50.781Z","hidden":false},{"_id":"6a8d0bb25add2537c32e9707","name":"Yanrui Miao","hidden":false},{"_id":"6a8d0bb25add2537c32e9708","user":{"_id":"676127cf11b19ea602bb202a","avatarUrl":"/avatars/dfd802a24bd63e509728159ebb1769f6.svg","isPro":false,"fullname":"Zhengxi Lu","user":"LZXzju","type":"user","name":"LZXzju"},"name":"Zhengxi Lu","status":"claimed_verified","statusLastChangedAt":"2026-08-27T08:45:04.860Z","hidden":false},{"_id":"6a8d0bb25add2537c32e9709","name":"Teng Pan","hidden":false},{"_id":"6a8d0bb25add2537c32e970a","name":"Yiwen Qiu","hidden":false},{"_id":"6a8d0bb25add2537c32e970b","name":"Hongxing Li","hidden":false},{"_id":"6a8d0bb25add2537c32e970c","name":"Peng Qiu","hidden":false},{"_id":"6a8d0bb25add2537c32e970d","name":"Ruiqing Zhang","hidden":false},{"_id":"6a8d0bb25add2537c32e970e","name":"Yongliang Shen","hidden":false}],"publishedAt":"2026-08-24T00:00:00.000Z","submittedOnDailyAt":"2026-08-27T00:00:00.000Z","title":"Agent-G^2: Gaussian Guidance for Agentic Reinforcement Learning","submittedOnDailyBy":{"_id":"676127cf11b19ea602bb202a","avatarUrl":"/avatars/dfd802a24bd63e509728159ebb1769f6.svg","isPro":false,"fullname":"Zhengxi Lu","user":"LZXzju","type":"user","name":"LZXzju"},"summary":"Hint-based reinforcement learning addresses reward sparsity in long-horizon agentic tasks by retaining a prefix of an expert trajectory before each rollout, letting the policy explore from a state closer to success. Its effectiveness hinges on the guidance depth: how much of the trajectory to keep. Existing methods treat this depth as a deterministic scalar. Scheduled approaches share one value across samples and ignore per-task heterogeneity; per-sample probing estimates it separately at the cost of extra rollouts. We find that useful guidance occupies a band of depths whose informativeness profile is approximately Gaussian around the band center, rather than concentrating at a single optimal point. We propose Agent-G^2, a Gaussian guidance framework that draws the depth per task from a Gaussian whose center and spread are estimated online from rollouts already collected for policy optimization, requiring no probe rollouts or learned depth predictor. The center combines a global baseline with per-cluster difficulty, and the spread tracks within-cluster variance. We evaluate Agent-G^2 on ALFWorld and WebShop on Qwen2.5-1.5B / 7B-Instruct. Agent-G^2 outperforms the strongest hint-based, hint-free, and Aux-RL baselines on ALFWorld by 2.3 / 3.9 / 7.4 points at under one-third the rollout cost of per-sample probing.","upvotes":17,"discussionId":"6a8d0bb25add2537c32e970f","projectPage":"https://zju-real.github.io/Agent-G2/","githubRepo":"https://github.com/ZJU-REAL/Agent-G2","githubRepoAddedBy":"user","ai_summary":"Agent-G² models hint depth as a Gaussian distribution estimated online from existing rollouts, improving reinforcement learning on long-horizon tasks without extra probing.","ai_keywords":["hint-based reinforcement learning","reward sparsity","expert trajectory","guidance depth","Gaussian guidance","Agent-G²","ALFWorld","WebShop"],"ai_summary_model":"thinkingmachines/Inkling-Small","githubStars":18,"organization":{"_id":"6345aadf5efccdc07f1365a5","name":"ZhejiangUniversity","fullname":"Zhejiang University","avatar":"https://www.gravatar.com/avatar/d1d414628877bec2958f95ad283c15e7?d=retro&size=100"}},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"677d3523a6918748bf81d0e9","avatarUrl":"/avatars/c9961ac54d089efb36db20c421c2bea2.svg","isPro":false,"fullname":"wangzixuan","user":"wangzx1210","type":"user"},{"_id":"66bc0ebd9c13cd4047a9bee0","avatarUrl":"/avatars/3bcf7d614ba6d6e975734eecc94fe037.svg","isPro":false,"fullname":"luka","user":"077lukamagic","type":"user"},{"_id":"698d83748068f1c868b8c0aa","avatarUrl":"/avatars/fb861fd898613f570478885512d046c1.svg","isPro":false,"fullname":"Yue","user":"WwwweweYue","type":"user"},{"_id":"6a0fe30ffb597becf7f6ea4f","avatarUrl":"/avatars/90c6a41e88f556a022566391ee1593fa.svg","isPro":false,"fullname":"Yue","user":"zhengwuyue","type":"user"},{"_id":"6a8f87a3056f4cd846f25ab2","avatarUrl":"/avatars/0afa5d4b4e23ac3e43cd7228aa17f07d.svg","isPro":false,"fullname":"Jinzhi Zhao","user":"Bonitoz","type":"user"},{"_id":"682b396401efbfd69d69a4ba","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/no-auth/8VxCrRbuyuqmRVtXl386n.png","isPro":false,"fullname":"Jiang","user":"RyanJiang0817","type":"user"},{"_id":"67543820c3af453d7b3e1d5e","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/67543820c3af453d7b3e1d5e/RbAZ9AQlxpy5E5Is-QN8b.jpeg","isPro":false,"fullname":"Dingming Li","user":"lidingm","type":"user"},{"_id":"676127cf11b19ea602bb202a","avatarUrl":"/avatars/dfd802a24bd63e509728159ebb1769f6.svg","isPro":false,"fullname":"Zhengxi Lu","user":"LZXzju","type":"user"},{"_id":"5e1058e9fcf41d740b69966d","avatarUrl":"/avatars/ce74839ba871f2b54313a670a233ba82.svg","isPro":false,"fullname":"Yongliang Shen","user":"tricktreat","type":"user"},{"_id":"67b970e414b1af2ce915c906","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/no-auth/2l7CwmVEs8NVZWcYgHk36.png","isPro":false,"fullname":"yiwen qiu","user":"qywMichelle","type":"user"},{"_id":"682c0fcebbe0c6fa323f531b","avatarUrl":"/avatars/953e6f5ce0361c2f9693bc0ca82787b7.svg","isPro":false,"fullname":"yy","user":"yuy07","type":"user"},{"_id":"68c7a335a7a25ec4cddf62be","avatarUrl":"/avatars/6cda5245a8a046aa455f9552888ed56d.svg","isPro":false,"fullname":"Kaiwen Zhang","user":"KeepKevin","type":"user"}],"acceptLanguages":["en"],"dailyPaperRank":0,"organization":{"_id":"6345aadf5efccdc07f1365a5","name":"ZhejiangUniversity","fullname":"Zhejiang University","avatar":"https://www.gravatar.com/avatar/d1d414628877bec2958f95ad283c15e7?d=retro&size=100"},"markdownContentUrl":"https://huggingface.co/buckets/huggingchat/papers-content/resolve/2608/2608.23318.md","query":{}}">

Papers

arxiv:2608.23318

Agent-G^2: Gaussian Guidance for Agentic Reinforcement Learning

Published on Aug 24

· Submitted by

Zhengxi Lu on Aug 27

Zhejiang University

Upvote

Authors:

Zixuan Wang ,

Zhengxi Lu ,

Abstract

Agent-G² models hint depth as a Gaussian distribution estimated online from existing rollouts, improving reinforcement learning on long-horizon tasks without extra probing.

Generated by thinkingmachines/Inkling-Small

Hint-based reinforcement learning addresses reward sparsity in long-horizon agentic tasks by retaining a prefix of an expert trajectory before each rollout, letting the policy explore from a state closer to success. Its effectiveness hinges on the guidance depth: how much of the trajectory to keep. Existing methods treat this depth as a deterministic scalar. Scheduled approaches share one value across samples and ignore per-task heterogeneity; per-sample probing estimates it separately at the cost of extra rollouts. We find that useful guidance occupies a band of depths whose informativeness profile is approximately Gaussian around the band center, rather than concentrating at a single optimal point. We propose Agent-G^2, a Gaussian guidance framework that draws the depth per task from a Gaussian whose center and spread are estimated online from rollouts already collected for policy optimization, requiring no probe rollouts or learned depth predictor. The center combines a global baseline with per-cluster difficulty, and the spread tracks within-cluster variance. We evaluate Agent-G^2 on ALFWorld and WebShop on Qwen2.5-1.5B / 7B-Instruct. Agent-G^2 outperforms the strongest hint-based, hint-free, and Aux-RL baselines on ALFWorld by 2.3 / 3.9 / 7.4 points at under one-third the rollout cost of per-sample probing.