Think Thrice Before You Speak: Dual knowledge-enhan...

Think Thrice Before You Speak: Dual knowledge-enhan... | AI Research

Key Takeaways

Think Thrice Before You Speak: Dual knowledge-enhanced Theory-of-Mind Reasoning for Persuasive Agents explores how AI can become more effective at persuasion...
Persuasive dialogue requires reasoning about others' latent mental states, a capability known as Theory of Mind (ToM).
To facilitate research on this task, we construct a large-scale annotated dataset, ToM-based Broad Persuasive Dialogues (ToM-BPD), capturing fine-grained mental states and corresponding persuasive strategies.
Experimental results demonstrate that Qwen3-8B equipped with TTBYS outperforms GPT-5 by 1.20%, 22.80%, and 16.97% in predicting desires, beliefs, and persuasive strategies, respectively.
Case studies further show that our approach enhances interpretability and consistency in reasoning.

Paper AbstractExpand

Persuasive dialogue requires reasoning about others' latent mental states, a capability known as Theory of Mind (ToM). However, due to reliance on simple prompting strategies and insufficient ToM knowledge, existing LLMs often fail to capture the intrinsic dependencies among mental states, leading to fragmented representations and unstable reasoning. To address these challenges, we introduce the ToM-based Persuasive Dialogue (ToM-PD) task, grounded in the Belief-Desire-Intention (BDI) framework, which explicitly models the sequential dependencies among mental states in multi-turn dialogues. To facilitate research on this task, we construct a large-scale annotated dataset, ToM-based Broad Persuasive Dialogues (ToM-BPD), capturing fine-grained mental states and corresponding persuasive strategies. We further propose Think Thrice Before You Speak (TTBYS), a knowledge-enhanced stepwise reasoning framework that leverages both explicit and implicit prior experiences to improve LLMs' inference of desires, beliefs, and persuasive strategies. Experimental results demonstrate that Qwen3-8B equipped with TTBYS outperforms GPT-5 by 1.20%, 22.80%, and 16.97% in predicting desires, beliefs, and persuasive strategies, respectively. Case studies further show that our approach enhances interpretability and consistency in reasoning.

Think Thrice Before You Speak: Dual knowledge-enhanced Theory-of-Mind Reasoning for Persuasive Agents explores how AI can become more effective at persuasion by better understanding human mental states. While current AI models can hold conversations, they often struggle to grasp the underlying beliefs and desires that drive a person's decisions. This research introduces a new framework that allows AI to "think" through these mental states step-by-step, leading to more consistent and human-like persuasive interactions.

Modeling the Human Mind

The researchers ground their work in the Belief-Desire-Intention (BDI) framework, a psychological model that explains how humans form actions. In this model, a person’s beliefs lead to desires, which then form intentions, eventually resulting in an action. While humans naturally follow this forward path, an AI trying to persuade someone must work backward: it observes a person’s words (actions) and must infer the intentions, desires, and beliefs that caused them. By explicitly modeling these dependencies, the AI can move away from fragmented, unstable reasoning and toward a more structured understanding of the person it is talking to.

The TTBYS Framework

To improve this reasoning, the authors developed the "Think Thrice Before You Speak" (TTBYS) framework. This approach acts as a knowledge-enhanced reasoning guide for Large Language Models (LLMs). Instead of jumping straight to a response, the AI uses TTBYS to retrieve both explicit and implicit prior experiences to help it analyze the user's mental state. By breaking the reasoning process into sequential steps—inferring intention, then desire, then belief, and finally the best persuasive strategy—the AI becomes more robust and capable of handling complex, multi-turn dialogues.

A New Benchmark for Persuasion

To support this research, the team created a large-scale dataset called ToM-based Broad Persuasive Dialogues (ToM-BPD). This dataset contains over 500 dialogues annotated with fine-grained information about the user's mental states and the specific persuasive strategies used by the AI. By using this data, the researchers were able to test how well different models perform in real-world scenarios.

Performance and Impact

The results show that equipping a model like Qwen3-8B with the TTBYS framework leads to significant improvements. In tests, this combination outperformed GPT-5 in predicting user desires, beliefs, and the most effective persuasive strategies. Beyond the numbers, the researchers found that their approach makes the AI’s reasoning more interpretable and consistent, providing a clearer path for developing AI agents that can engage in more meaningful and effective social interactions.