Researchers Use AI Agents to Generate Persuasive Dialogue Training Data at Scale

Researchers have developed a novel system for generating large volumes of persuasive dialogue training data using coordinated AI agents in self-play scenarios, avoiding the expense and complexity of manual annotation.

How MADS Works#

The framework, called MADS (Multi-Agent Dialogue Simulation), orchestrates three specialized agents working in tandem. User Agents mimic varied customer personas by incorporating personality dimensions drawn from astrology (Zodiac signs) and psychological profiling (MBTI types), creating naturalistic behavioral diversity. A Dialog Agent manages the persuasion task itself, employing targeted strategies to sway users. Finally, an Optimization Agent assesses dialogue quality and iteratively improves the outcomes, refining the synthetic conversations through feedback loops.

Real-World Impact#

When tested in an actual marketing deployment, the approach delivered tangible business results. Small language models trained on MADS-generated dialogues achieved a 22.4% lift in conversion rates, rising from 1.83% to 2.24% of organic traffic. This performance gain demonstrates that synthetic, agent-driven training data can meaningfully enhance persuasion capabilities in production environments.

Why This Matters#

Generating persuasive dialogue at scale has long posed operational challenges: scarcity of real user interaction data, difficulty establishing baseline performance early in product development, and inefficient prompt engineering. By automating data synthesis through agent roleplay, MADS sidesteps these bottlenecks and reduces labor costs. The system also incorporates Chain-of-Attitude (CoA) modeling to evaluate persuasiveness from the user's perspective and employs dedicated LLMs as independent judges of dialogue quality, adding rigor to the evaluation process.