Check out our new work, <strong>HASP</strong>, where we turn reusable agent skills from passive text prompts into executable program functions that can actively intervene in the agent loop for stronger inference-time control, post-training supervision, and self-improvement!</p>\n","updatedAt":"2026-05-20T15:47:00.253Z","author":{"_id":"646474126d621be0dd08d6d4","avatarUrl":"/avatars/5041573e603225e9de3b689efd1dd499.svg","fullname":"Liu Hongjun","name":"Jan150000","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"isUserFollowing":false}},"numEdits":1,"identifiedLanguage":{"language":"en","probability":0.8683958053588867},"editors":["Jan150000"],"editorAvatarUrls":["/avatars/5041573e603225e9de3b689efd1dd499.svg"],"reactions":[],"isReport":false}},{"id":"6a0de1e995b817c6dbabc123","author":{"_id":"6333fa238003819e16ae117e","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6333fa238003819e16ae117e/sTliVSzXNJNLIComJP99G.jpeg","fullname":"liu","name":"carton","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":1,"isUserFollowing":false},"createdAt":"2026-05-20T16:31:37.000Z","type":"comment","data":{"edited":false,"hidden":false,"latest":{"raw":"Fantastic work! Thumbs up!👍","html":"<p>Fantastic work! Thumbs up!👍</p>\n","updatedAt":"2026-05-20T16:31:37.983Z","author":{"_id":"6333fa238003819e16ae117e","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6333fa238003819e16ae117e/sTliVSzXNJNLIComJP99G.jpeg","fullname":"liu","name":"carton","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":1,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.8346078991889954},"editors":["carton"],"editorAvatarUrls":["https://cdn-avatars.huggingface.co/v1/production/uploads/6333fa238003819e16ae117e/sTliVSzXNJNLIComJP99G.jpeg"],"reactions":[{"reaction":"🚀","users":["Jan150000"],"count":1}],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2605.17734","authors":[{"_id":"6a0c948765eb30f20d962961","user":{"_id":"646474126d621be0dd08d6d4","avatarUrl":"/avatars/5041573e603225e9de3b689efd1dd499.svg","isPro":false,"fullname":"Liu Hongjun","user":"Jan150000","type":"user","name":"Jan150000"},"name":"Hongjun Liu","status":"claimed_verified","statusLastChangedAt":"2026-05-20T17:12:24.382Z","hidden":false},{"_id":"6a0c948765eb30f20d962962","name":"Yifei Ming","hidden":false},{"_id":"6a0c948765eb30f20d962963","name":"Shafiq Joty","hidden":false},{"_id":"6a0c948765eb30f20d962964","name":"Chen Zhao","hidden":false}],"publishedAt":"2026-05-18T00:00:00.000Z","submittedOnDailyAt":"2026-05-20T00:00:00.000Z","title":"Harnessing LLM Agents with Skill Programs","submittedOnDailyBy":{"_id":"646474126d621be0dd08d6d4","avatarUrl":"/avatars/5041573e603225e9de3b689efd1dd499.svg","isPro":false,"fullname":"Liu Hongjun","user":"Jan150000","type":"user","name":"Jan150000"},"summary":"Equipping LLM agents with reusable skills derived from past experience has become a popular and successful approach for tackling complex and long-horizon tasks. However, such lessons are often encoded as textual guidance that remains largely advisory, lacking explicit mechanisms for when and how to intervene in the agent loop. To bridge the gap, we introduce HASP(Harnessing LLM Agents with Skill Programs), a new framework that upgrades skills into executable Program Functions (PFs). Rather than offering passive advice, PFs act as executable guardrails that activate on failure-prone states and modify the next action or inject corrective context. HASP is highly modular: it can be applied at inference time for direct agent-loop intervention, during post-training to provide structured supervision, or for self-improvement by evolving validated, teacher-reviewed PFs. Empirically, HASP drives substantial gains compared to both training-free and training-based methods on web-search, math reasoning, and coding tasks. For example, on web-search reasoning, inference-time PFs alone improve the average performance by 25% compared to (multi-loop) ReAct Agent, while post-training and controlled evolution achieve a 30.4% gain over Search-R1. To provide deeper insights into HASP, our mechanism analysis reveals how PFs trigger and intervene, how skills are internalized, and the requirement for stable skill library evolution.","upvotes":27,"discussionId":"6a0c948765eb30f20d962965","ai_summary":"HASP introduces executable program functions that serve as active guardrails for LLM agents, enabling direct intervention in agent loops and improving performance across complex tasks.","ai_keywords":["LLM agents","skill programs","Program Functions","executable guardrails","agent loop intervention","post-training","self-improvement","web-search reasoning","math reasoning","coding tasks","ReAct Agent","Search-R1"],"organization":{"_id":"691d8e884bbe8df0d99462e2","name":"newyorkuniversity","fullname":"New York University","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/68e396f2b5bb631e9b2fac9a/orNHmPzOQf2_F5UgXPsu5.png"}},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"646474126d621be0dd08d6d4","avatarUrl":"/avatars/5041573e603225e9de3b689efd1dd499.svg","isPro":false,"fullname":"Liu Hongjun","user":"Jan150000","type":"user"},{"_id":"67dc3e58b505373398cf4635","avatarUrl":"/avatars/130369ded932c974b5123416e7e5a249.svg","isPro":false,"fullname":"Yuhui Wang","user":"Michael109","type":"user"},{"_id":"66229ec8c8920ec35141ae86","avatarUrl":"/avatars/0c64e68e836c3477713385f9a0a70a89.svg","isPro":false,"fullname":"Long Chen","user":"KuroSkyyy","type":"user"},{"_id":"671ffc50dc8f614649f15e25","avatarUrl":"/avatars/4dbda057baf68d11e56efa88424a1e7f.svg","isPro":false,"fullname":"Renyun Li","user":"renyunli0116","type":"user"},{"_id":"6333fa238003819e16ae117e","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6333fa238003819e16ae117e/sTliVSzXNJNLIComJP99G.jpeg","isPro":false,"fullname":"liu","user":"carton","type":"user"},{"_id":"616745f893f6837511d9ae25","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/616745f893f6837511d9ae25/H0JIGDL5VOu29zVd4KiKB.jpeg","isPro":false,"fullname":"Kaiser Sun","user":"KaiserWhoLearns","type":"user"},{"_id":"68ea4b411b306d9020c31c2c","avatarUrl":"/avatars/43f9b4cccbf67de43b7dd5edc5f1392b.svg","isPro":false,"fullname":"Peng Wang","user":"93393pwa","type":"user"},{"_id":"6a0de515d9650f0eef5a3ce5","avatarUrl":"/avatars/a65d3a5d6a6dd7b165d904ac84d7191f.svg","isPro":false,"fullname":"Liu","user":"HJ15000","type":"user"},{"_id":"655cfc0fd74f417795ac0708","avatarUrl":"/avatars/0661d67b47d5f5ae9359a881efde79f7.svg","isPro":false,"fullname":"LUZEXIN","user":"LUZEXIN","type":"user"},{"_id":"648d1b77d69d6e09bcff14fd","avatarUrl":"/avatars/f701ce5e7f00a3da6fb067f7e9220b8e.svg","isPro":false,"fullname":"Ma Jiajian","user":"idiotbot2","type":"user"},{"_id":"64f98808b92d48fbea02c852","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/64f98808b92d48fbea02c852/PJZ2zlZrYW0ojaj3dUKRW.jpeg","isPro":false,"fullname":"James Luo","user":"horizonariadust","type":"user"},{"_id":"648749094dea003c6dae810f","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/648749094dea003c6dae810f/gHUHSBt1zrt8wjO1YwTNu.jpeg","isPro":false,"fullname":"Shrey Pandit","user":"SP2001","type":"user"}],"acceptLanguages":["en"],"dailyPaperRank":0,"organization":{"_id":"691d8e884bbe8df0d99462e2","name":"newyorkuniversity","fullname":"New York University","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/68e396f2b5bb631e9b2fac9a/orNHmPzOQf2_F5UgXPsu5.png"},"markdownContentUrl":"https://huggingface.co/buckets/huggingchat/papers-content/resolve/2605/2605.17734.md"}">
Harnessing LLM Agents with Skill Programs
Abstract
HASP introduces executable program functions that serve as active guardrails for LLM agents, enabling direct intervention in agent loops and improving performance across complex tasks.
AI-generated summary
Equipping LLM agents with reusable skills derived from past experience has become a popular and successful approach for tackling complex and long-horizon tasks. However, such lessons are often encoded as textual guidance that remains largely advisory, lacking explicit mechanisms for when and how to intervene in the agent loop. To bridge the gap, we introduce HASP(Harnessing LLM Agents with Skill Programs), a new framework that upgrades skills into executable Program Functions (PFs). Rather than offering passive advice, PFs act as executable guardrails that activate on failure-prone states and modify the next action or inject corrective context. HASP is highly modular: it can be applied at inference time for direct agent-loop intervention, during post-training to provide structured supervision, or for self-improvement by evolving validated, teacher-reviewed PFs. Empirically, HASP drives substantial gains compared to both training-free and training-based methods on web-search, math reasoning, and coding tasks. For example, on web-search reasoning, inference-time PFs alone improve the average performance by 25% compared to (multi-loop) ReAct Agent, while post-training and controlled evolution achieve a 30.4% gain over Search-R1. To provide deeper insights into HASP, our mechanism analysis reveals how PFs trigger and intervene, how skills are internalized, and the requirement for stable skill library evolution.
Community
Check out our new work, HASP, where we turn reusable agent skills from passive text prompts into executable program functions that can actively intervene in the agent loop for stronger inference-time control, post-training supervision, and self-improvement!
Fantastic work! Thumbs up!👍
Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.
Tap or paste here to upload images
Cite arxiv.org/abs/2605.17734 in a model README.md to link it from this page.
Cite arxiv.org/abs/2605.17734 in a dataset README.md to link it from this page.
Cite arxiv.org/abs/2605.17734 in a Space README.md to link it from this page.
Discussion (0)
Sign in to join the discussion. Free account, 30 seconds — email code or GitHub.
Sign in →No comments yet. Sign in and be the first to say something.