This paper proposes Multi-Agent Dialogue Simulation (MADS), a scalable framework for generating persuasive multi-turn conversations through agent self-play. MADS utilizes three collaborative agents: User Agents, which simulate various persona-based behaviors using personality indicators such as zodiac signs and Myers-Briggs Types; Dialog Agents, which execute task-oriented persuasive strategies; and Optimization Agents, which evaluate and improve conversation outcomes. We validate its effectiveness by modeling users' Chain-of-Attitude (CoA) and evaluating the persuasiveness of a dedicated LLM. This approach enables low-cost training data generation without human annotation, addressing key industry challenges such as lack of user data, difficulties in cold-start evaluation, and inefficient prompts. When applied to a real-world marketing scenario, MADS significantly enhanced the persuasiveness of a small LLM, increasing organic traffic conversion rates by 22.4% (from 1.83% to 2.24%), demonstrating clear business value.
