ToMAP: Training Opponent-Aware LLM Persuaders with Theory of Mind
Researchers have developed ToMAP, a new approach to train large language models for persuasion by incorporating Theory of Mind (ToM) modules. These modules enhance the model's ability to understand and adapt to an opponent's mental state and potential objections. Experiments show that the 3B parameter ToMAP model outperforms significantly larger models like GPT-4o, achieving a 39.4% relative gain in effectiveness across various corpora. ToMAP's opponent-aware capabilities lead to more diverse, logical, and less repetitive arguments, making it suitable for extended conversational interactions. AI
IMPACT Enhances LLM capabilities in strategic communication and negotiation by improving opponent modeling.