A simple problem we ran into: prompt optimization can improve a multi-agent system while quietly breaking its routing, formatting, or termination logic. We separate control flow into typed program objects and only optimize the natural-language data flow, which gives 100% eventual protocol validity while still improving task performance.</p>\n","updatedAt":"2026-09-02T02:00:26.788Z","author":{"_id":"63081e15a670ed10f9d44229","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/63081e15a670ed10f9d44229/w1b9uq-9774bMMgJbSPsS.jpeg","fullname":"Yuntian Deng","name":"yuntian-deng","type":"user","isPro":true,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":277,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.8627097010612488},"editors":["yuntian-deng"],"editorAvatarUrls":["https://cdn-avatars.huggingface.co/v1/production/uploads/63081e15a670ed10f9d44229/w1b9uq-9774bMMgJbSPsS.jpeg"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2609.00621","authors":[{"_id":"6a977ccefe3c2f89286c38b6","name":"Wentao Zhang","hidden":false},{"_id":"6a977ccefe3c2f89286c38b7","name":"Syed Shariyar Murtaza","hidden":false},{"_id":"6a977ccefe3c2f89286c38b8","name":"Junaid Ahmad Bhatti","hidden":false},{"_id":"6a977ccefe3c2f89286c38b9","name":"Utkarsh Soni","hidden":false},{"_id":"6a977ccefe3c2f89286c38ba","name":"Yifan Nie","hidden":false},{"_id":"6a977ccefe3c2f89286c38bb","name":"Eugene Wen","hidden":false},{"_id":"6a977ccefe3c2f89286c38bc","name":"Yuntian Deng","hidden":false}],"publishedAt":"2026-09-01T00:00:00.000Z","submittedOnDailyAt":"2026-09-02T00:00:00.000Z","title":"Control-Data Flow Separation: Stable Prompt Optimization in Multi-Agent LLMs","submittedOnDailyBy":{"_id":"63081e15a670ed10f9d44229","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/63081e15a670ed10f9d44229/w1b9uq-9774bMMgJbSPsS.jpeg","isPro":true,"fullname":"Yuntian Deng","user":"yuntian-deng","type":"user","name":"yuntian-deng"},"summary":"Prompt optimization can improve multi-agent LLM systems, but the prompts being optimized often serve two entangled roles: generating task-relevant content and specifying execution-critical protocols, such as message routing, output formatting, and termination signals, on which the underlying code relies. As a result, a prompt edit intended to improve content generation can inadvertently corrupt the protocol and cause the entire agent pipeline to fail. Our key observation is that these two roles have different representations: execution protocols are typically structured, while task-relevant content is usually expressed in unstructured language. Based on this, we propose control-data flow separation, where execution-critical control is represented as typed, validated program objects, while task-relevant language remains the optimizable data flow for agent communication. This design allows optimizers to improve multi-agent behavior without exposing the routing or formatting interface to prompt drift. Across synthetic reasoning, collaborative review generation, and insurance rating workflows, our framework empirically achieves 100% eventual protocol validity while consistently improving task performance.","upvotes":5,"discussionId":"6a977ccefe3c2f89286c38bd","githubRepo":"https://github.com/yuntian-group/cdsep","githubRepoAddedBy":"user","ai_summary":"The framework separates structured execution protocols from optimizable language content to prevent prompt optimization from corrupting multi-agent pipelines.","ai_keywords":["multi-agent LLM systems","prompt optimization","control-data flow separation","typed validated program objects","protocol validity"],"ai_summary_model":"thinkingmachines/Inkling-Small","githubStars":1,"organization":{"_id":"6230c3ced93e84e2338765f3","name":"UWaterloo","fullname":"University of Waterloo","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/1647363004040-6230c34bacb94a81eec91ed5.png"}},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"66387baa9ff6e77b0042652c","avatarUrl":"/avatars/c365bc3e1963dffe338885d62ff6bc20.svg","isPro":false,"fullname":"Wentao Zhang","user":"wtzhang-nlp","type":"user"},{"_id":"63081e15a670ed10f9d44229","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/63081e15a670ed10f9d44229/w1b9uq-9774bMMgJbSPsS.jpeg","isPro":true,"fullname":"Yuntian Deng","user":"yuntian-deng","type":"user"},{"_id":"63ac5701c21e60a3e9b58aa7","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/63ac5701c21e60a3e9b58aa7/g6EX7diOpuA94R2ab-rZC.png","isPro":true,"fullname":"Dipankar Sarkar","user":"dipankarsarkar","type":"user"},{"_id":"654ed6406c46a91f120443ba","avatarUrl":"/avatars/4e0c2c7e56d648ed24c3b079acd237da.svg","isPro":false,"fullname":"Boren","user":"borent","type":"user"},{"_id":"66a017ead2b13e3f33d03725","avatarUrl":"/avatars/17c995158db3365f066f26cd4a3bbe29.svg","isPro":false,"fullname":"Boren Hong","user":"hongb2","type":"user"}],"acceptLanguages":["en"],"dailyPaperRank":0,"organization":{"_id":"6230c3ced93e84e2338765f3","name":"UWaterloo","fullname":"University of Waterloo","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/1647363004040-6230c34bacb94a81eec91ed5.png"},"markdownContentUrl":"https://huggingface.co/buckets/huggingchat/papers-content/resolve/2609/2609.00621.md","query":{}}">

Get this paper in your agent:

Don't have the latest CLI?

Cite arxiv.org/abs/2609.00621 in a model README.md to link it from this page.

Cite arxiv.org/abs/2609.00621 in a dataset README.md to link it from this page.

Cite arxiv.org/abs/2609.00621 in a Space README.md to link it from this page.

Add this paper to a collection to link it from this page.