Navigating Attitude Transitions: A Train-Free Strategy Planner for Multi-Turn Persuasion

ACL ARR 2026 January Submission2588 Authors

Keywords: Persuasion, Attitude-Aware, Strategy Planner, Attitude Transitions, Train-free Planner

Abstract: Multi-turn persuasion is often formulated as an end-to-end text generation problem that focuses on the persuadee’s final attitude, leaving intermediate attitude transitions implicit. As a result, strategic decisions are implicitly entangled with language realization, leading to unstructured strategy use and limited controllability across dialogue turns. In this work, we reformulate multi-turn persuasion as a process of navigating latent attitude transitions through explicit turn-level strategy planning. We propose an attitude-aware framework that decouples strategy selection from response generation. At each turn, the persuadee’s latent attitude is inferred, a persuasion strategy is selected by a dedicated planner, and a language model generates a strategy-conditioned response. We further introduce a train-free strategy planner grounded in empirical attitude–strategy transition statistics, enabling explicit and stable strategy selection without additional training. Experiments demonstrate that our framework consistently outperforms end-to-end persuasion and attitude-conditioned generation without explicit planning, achieving up to 95.4% acceptance rates with fewer dialogue turns. Further analysis shows that different planners exhibit distinct strategy selection patterns, resulting in different persuasion dynamics. Notably, the proposed train-free planner matches or even surpasses LLM-based planners in several settings, particularly when LLM-based strategy selection is unstable, highlighting the robustness and reliability benefits of explicit strategy planning for multi-turn persuasion.

Paper Type: Long

Research Area: Dialogue and Interactive Systems

Research Area Keywords: task-oriented, applications, dialogue state tracking

Contribution Types: Model analysis & interpretability, NLP engineering experiment, Data resources, Theory

Languages Studied: English

Submission Number: 2588