Control-Data Flow Separation: Stable Prompt Optimization in Multi-Agent LLMs
The paper’s fix is to keep agent protocols out of the prompts being optimized.
It argues that multi-agent prompt optimization can break systems when edits alter routing, formatting, or termination rules that code depends on. The proposed control-data flow separation moves those execution rules into typed, validated program objects, leaving only task language exposed to optimization. In tests across reasoning, review generation, and insurance rating workflows, the framework reached 100% eventual protocol validity while improving task performance. HF Daily Papers' note
It argues that multi-agent prompt optimization can break systems when edits alter routing, formatting, or termination rules that code depends on. The proposed control-data flow separation moves those execution rules into typed, validated program objects, leaving only task language exposed to optimization. In tests across reasoning, review generation, and insurance rating workflows, the framework reached 100% eventual protocol validity while improving task performance. HF Daily Papers' note
score 4