Communication Policy Evolution Enhances Proactive LLM Agent Interaction.

As Large Language Model (LLM) agents become more autonomous, a persistent challenge is the information gap between users and agents, compounded by the cost of communication and often identical user preferences. This paper explores how agents should communicate across different modalities by formalizing "Communication Policy" and evaluating textual and UI-based policies in various environments, with different personas and model combinations. The study establishes two complementary settings, User-Agent and Planner-Executor, to build information asymmetry for proactive agents. Experimental findings highlight the distinct strengths of different interaction channels: text-based communication often aids task performance, while structured UI improves agent response quality and adherence to persona. Motivated by these observations, a hybrid communication method is proposed to combine these advantages. Furthermore, the paper introduces Communication Policy Evolution (CPE), a self-evolution framework designed to refine communication policies. CPE achieves this through rollout and prompt-level evolving, demonstrating that prompt refinement alone can lead to the best task success across multiple settings, without requiring any modifications to the underlying LLM. This identifies communication behavior as a critical, yet often overlooked, design dimension for LLM agents.