ToMAP: Endowing Large Models with 'Mind Reading' to Create Smarter AI Persuaders}

Persuasion is a process that influences others' beliefs, attitudes, and behaviors, widely present in human society. As a complex and common form of communication, it also serves as a benchmark for the capabilities of increasingly powerful large language models.

Top-tier large models can generate coherent persuasive passages and even mimic human-like interactions on platforms like Reddit. However, their lack of mental perception limits their ability to enhance persuasion further.

Successful persuasion requires not only clear arguments but also precise insight into the other party’s stance and thought process. This insight, known as "Theory of Mind" (ToM), involves recognizing that others have independent thoughts, beliefs, and motivations, and reasoning based on this understanding. Humans possess this innate cognitive ability, but large models often lack it, leading to two major shortcomings:

• Models tend to focus only on core arguments without generating new perspectives based on the connections between points.
• Models often repeat their own viewpoints without adjusting strategies based on the opponent’s attitude changes.

To address this, researchers at UIUC proposed ToMAP (Theory of Mind Augmented Persuader), a new persuasion model that incorporates ToM mechanisms, enabling AI to "put itself in the opponent’s shoes" for more personalized, flexible, and logical persuasion.

• Paper Title: ToMAP: Training Opponent-Aware LLM Persuaders with Theory of Mind
• Paper Link: https://arxiv.org/pdf/2505.22961
• Open Source Code: https://github.com/ulab-uiuc/ToMAP

ToMAP: Knowing Yourself and Your Enemy, Winning Every Battle

ToMAP innovatively introduces two major mental modules into the persuader framework: Rebuttal Predictor and Attitude Predictor.

Rebuttal Predictor simulates humans’ proactive anticipation of potential objections during persuasion. The study found that large models possess this ability, which can be activated through prompt design. Qualitative and quantitative analyses show that the rebuttals generated by the model are semantically similar to the actual objections of persuaded individuals, giving the AI a "first-mover" advantage to proactively address doubts. For example, in advocating a "vegetarian recipe," the predictor can identify reasons like "cooking difficulty" and "bad taste" as objections, forming a complex argument network around the core point.

Recognizing rebuttals alone isn’t enough to capture complex attitude shifts. Therefore, the Attitude Predictor further assesses the opponent’s stance towards these rebuttals—whether they agree, are neutral, or persuaded. This module uses dialogue history and arguments as input, employing BGE-M3 text encoder and multi-layer perceptron (MLP) classifier to dynamically estimate the opponent’s attitude, enabling targeted argumentation.

Experiments show that the predictor outperforms direct large model reasoning in predicting attitudes. For instance, in a dialogue, the opponent recognizes the health benefits of vegetarianism but mentions not "enjoying" it, indicating a possible reservation about taste, which guides the next persuasion focus.

The integration of these two modules allows the persuader to access richer information: predicting potential objections and dynamically evaluating the opponent’s mental state. This facilitates more diverse and targeted dialogues, effectively influencing opinions.

However, large language models may not fully utilize this information. To maximize these advantages, ToMAP employs Reinforcement Learning (RL), training the model through numerous dialogues. The model is rewarded based on a "persuasiveness score," reflecting attitude changes after each interaction. Additional signals like format rewards, repetition penalties, and over-length penalties help generate coherent, persuasive conversations.

Analysis: Strategic Planning for Victory

Systematic testing on various datasets and opponent models shows that the ToMAP model significantly outperforms baseline models and RL versions without mental modules. Notably, despite having only 3 billion parameters, ToMAP surpasses larger models like GPT-4o and DeepSeek-R1, demonstrating that with proper training and module design, smaller models can achieve remarkable persuasion capabilities.

Reviewing the training trajectory of ToMAP reveals that as persuasion rewards increase, the model maintains low repetition penalties, indicating effective use of mental modules to enhance output diversity. Its reasoning depth also increases, showing strategic thinking capabilities. The model favors rational, targeted strategies over emotional or authoritative tactics, which boosts persuasion.

Furthermore, ToMAP demonstrates stable improvement in persuasion over long dialogues, unlike baseline and standard RL models, which tend to plateau or decline after a few rounds. This highlights its strong strategic adaptability and argument diversity.

Conclusion: Injecting "Human-like Cognition" into AI

This research introduces ToMAP, a framework combining theory of mind with AI persuasion, addressing the lack of opponent modeling and strategic flexibility in current large language models. By simulating human anticipatory reasoning and attitude perception, ToMAP makes AI more perceptive and adaptable in persuasion tasks. The reinforcement learning mechanism encourages diverse, well-structured, and logical argumentation.

Beyond improving persuasion, ToMAP surpasses strong baselines across multiple datasets and model combinations, marking a significant step toward AI with "social cognition." Its ability to actively understand the opponent’s cognitive structure and attitude tendencies demonstrates initial "social awareness," making language models more humanized and strategic in complex interactions.

In summary, ToMAP is not only an effective training framework for persuasive AI but also an innovative step toward developing models with "human-like thinking," providing a solid foundation for trustworthy, flexible AI communication systems.