Teaching AI to persuade by reading minds it can't see

Persuading someone requires reading between the lines: understanding unstated beliefs, desires, and resistance. This paper builds a multi-agent system that combines perception analysis, mental-state inference, strategy selection, and memory to generate persuasive dialogue in complex scenarios where the other person doesn't spell out their position. A key innovation is a "meta-cognitive configurator" that picks the right persuasion strategy upfront, anchoring all downstream reasoning. The system beats baseline LLM approaches on success rate across multiple domains.