MADS: Multi-Agent Dialogue Simulation for Diverse Persuasion Data Generation

TL;DR

MADS presents a scalable framework for generating persuasive multi-turn dialogues via multi-agent self-play, leveraging persona-driven User Agents and an Optimization Agent to produce high-quality training data without human annotation. The methodology centers on Chain-of-Attitude (CoA), a 16-state attitude model with a first-order Markov transition and entropy-based diversity measures, enabling richer behavioral trajectories and better coverage of persuasive strategies. Through a self-optimizing loop and domain-specific data augmentation, MADS improves the persuasive capabilities of small LLMs, with validated gains on MMP and P4G benchmarks and tangible business impact in a real-world marketing scenario, including a notable uplift in organic conversion rate. Overall, the work demonstrates that symbolic personality cues and automated prompt evolution can significantly enhance data diversity and downstream performance in persuasion-focused dialogue systems, offering practical benefits for cold-start deployments.

Abstract

We propose MADS (Multi-Agent Dialogue Simulation), a scalable framework for generating persuasive multi-turn dialogues via agent self-play. MADS employs three coordinated agents: User Agents designed to simulate diverse persona-driven behaviors by leveraging personality signifiers such as Zodiac Signs and MBTI types, a Dialog Agent executing task-oriented persuasion strategies and an Optimization Agent evaluating and refining dialogue outcomes. We further validate its effectiveness through users' Chain-of-Attitude (CoA) modeling and dedicated LLMs' persuasion assessment. This approach enables low-cost generation of training data without human annotation, addressing key industry challenges such as lack of user data, cold-start evaluation difficulties, and prompt inefficiency. Applied to a real-world marketing scenario, MADS significantly improved the persuasion capacity of small LLMs, increasing the organic traffic conversion rate by 22.4% (from 1.83% to 2.24%) , demonstrating clear business value.

Figures (5)

• Figure 1: Architecture of the MADS Framework
• Figure 2: Visualization of attitude transition matrices across dialogue collections of different user agent persona groups, (a) $D_{base}$, (b) $D_{sign}$, (c) $D_{mbti}$, (d) $D_{busi}$
• Figure 3: P4G Results Distribution of persuasive strategy types used across different user persona groups
• Figure :
• Figure :