TTBYS: A Dual Knowledge-Enhanced Theory of Mind Reasoning Framework for Persuasive Dialogue

This article introduces the TTBYS (Think Three Times Before You Speak) framework, which enhances the Theory of Mind reasoning ability of large language models through explicit and implicit prior knowledge, enabling more stable persuasive dialogue generation under the Belief-Desire-Intention (BDI) framework. The framework performs excellently on the ToM-BPD dataset; using the Qwen3-8B model, it surpasses GPT-5 in key metrics such as desire prediction, belief reasoning, and persuasive strategy selection, providing new ideas for building socially intelligent AI systems.

Persuasive dialogue is an advanced human cognitive activity that requires accurate reasoning about others' beliefs, desires, and intentions (Theory of Mind, ToM). Current large language models face challenges: simple prompting strategies and limited ToM knowledge make it difficult to capture dependencies between mental states, leading to fragmented representations and unstable reasoning results.

ToM-PD Task Framework

The study proposes the Theory of Mind-based Persuasive Dialogue task (ToM-PD), built on the BDI framework, which explicitly models the sequential dependencies of mental states in multi-turn dialogues and provides a structured task definition.

ToM-BPD Dataset

A large-scale annotated dataset ToM-BPD is constructed with the following features:

1. Fine-grained mental state annotations (beliefs, desires, intentions);
2. Corresponding persuasive strategy annotations, establishing mappings from mental states to strategies;
3. Multi-turn dialogue structure to capture dynamic changes in mental states.

TTBYS Framework: Think Three Times Before You Speak

The core consists of three思考 steps before generating a response:

1. Desire Reasoning: Use prior knowledge to infer the most urgent desires/needs of the dialogue partner;
2. Belief Reasoning: Based on desires, infer the partner's belief state regarding the topic (belief, doubt, openness);
3. Strategy Selection: Synthesize the results of the previous two steps to select effective persuasive strategies, drawing on historical experience to avoid repeated failures.

Dual Knowledge Enhancement

1. Explicit Prior Knowledge: Psychological theories, persuasion research findings, and manually annotated case libraries, injected into the model in a structured manner;
2. Implicit Experiential Knowledge: Retrieve similar historical dialogue cases to supplement the limitations of explicit theories and enhance flexibility.

Experimental Comparison

On the ToM-BPD dataset, Qwen3-8B+TTBYS outperforms GPT-5:

• Desire prediction accuracy increased by 1.20%;
• Belief reasoning accuracy increased by 22.80%;
• Persuasive strategy selection accuracy increased by 16.97%.

Ablation Experiments

Removing either explicit knowledge or implicit experience leads to a significant drop in performance, proving their complementarity.

Case Studies

• Visualization of reasoning process: Step-by-step reasoning outputs intermediate steps, making decisions transparent and interpretable;
• Cross-turn consistency: More stable understanding of mental states in multi-turn dialogues, leading to more coherent conversations.

1. Theory of Mind Engineering: Transform abstract ToM concepts into computable components, providing methodological references for social reasoning applications;
2. Knowledge Enhancement Paradigm: The dual knowledge enhancement approach of explicit theory + implicit experience can be extended to fields such as medical diagnosis, legal consultation, and educational tutoring;
3. Potential of Small Models: Medium-sized models (e.g., Qwen3-8B) can outperform ultra-large models in specific tasks through architectural design and knowledge injection.

1. Dataset: Although ToM-BPD is large-scale, its domain coverage and language diversity need to be expanded;
2. Efficiency: Three-step sequential reasoning increases computational overhead; reasoning speed needs to be optimized for practical deployment;
3. Ethics: Persuasion technology is a double-edged sword; it needs to be applied carefully to avoid manipulation or deception.

The TTBYS framework integrates Theory of Mind, BDI modeling, and dual knowledge enhancement, providing a new solution for persuasive dialogue. Its "Think Three Times Before You Speak" design improves performance and enhances the interpretability and consistency of reasoning. This work takes an important step forward in the development of socially intelligent AI systems, while also reminding us to pay attention to ethical boundaries and social impacts.