Planning Persuasion, Not Utterances: Profile-Conditioned Open-Loop Search for Dialogue Strategy

Persuasive dialogue systems are increasingly important for prosocial communication, education, and support-oriented interaction, but they still struggle to make profile-sensitive decisions when user reactions are uncertain and persuasive outcomes are delayed. This paper aims to enable stable long-horizon persuasive planning that adapts to user profiles while avoiding brittle text-level search and costly prompt-based self-evaluation. We propose PMCTS-PD, a profile-conditioned open-loop tree search framework that plans over compact dialogue-act prefixes rather than full utterances, thereby separating strategic decision making from surface realization. Each search node aggregates multiple instantiated dialogue histories, while a lightweight profile-conditioned Value-LLM predicts normalized expected donation to guide tree search, value backup, and value-guided best-of-K response generation. On PersuasionForGood, PMCTS-PD improves over strong LLM and dialogue-planning baselines, achieving higher BLEU-4, embedding similarity, diversity, predicted donation, and end-to-end success rate. These results show that planning persuasion at the strategy level, supported by a learned profile-aware outcome signal, offers a practical and cost-effective direction for robust persuasive dialogue systems.