Title: ToMAP: Training Opponent-Aware LLM Persuaders with Theory of Mind

Abstract:

While Large Language Models (LLMs) demonstrate significant promise in the realm of persuasion, current methodologies for training persuasive LLMs remain in their nascent stages. Humans excel at proactively and dynamically modeling their opponents' thoughts and opinions; however, contemporary LLMs lack this Theory of Mind (ToM) reasoning capability, which restricts both the diversity of their arguments and their awareness of the opponent. To overcome this deficiency, we propose Theory of Mind Augmented Persuader (ToMAP), a novel framework designed to construct more adaptable persuader agents by integrating two distinct Theory of Mind modules. These modules significantly improve the persuader’s ability to analyze and perceive the mental state of the opponent.

The process begins by prompting the persuader to anticipate potential objections to the target central claim. Subsequently, a text encoder combined with a trained Multi-Layer Perceptron (MLP) classifier is employed to predict the opponent’s current stance regarding these counterarguments. Through a carefully engineered reinforcement learning schema, the persuader acquires the skill to analyze opponent-specific data and leverage it to formulate more compelling arguments.

Experimental results indicate that ToMAP, despite having only 3 billion parameters, surpasses significantly larger baseline models, such as GPT-4o, achieving a relative performance gain of 39.4% across various persuadee models and diverse datasets. Notably, the training process fosters complex reasoning chains and minimizes repetition, thereby yielding more varied and effective arguments. Furthermore, ToMAP’s opponent-aware architecture makes it particularly well-suited for extended conversations, allowing it to deploy more logical and context-sensitive strategies. These findings validate the effectiveness of our approach and emphasize its potential for advancing the development of highly persuasive language agents. The code is publicly available at: https://github.com/ulab-uiuc/ToMAP.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC