Abstract
Persuasion, as one of the crucial abilities in human communication, has garnered extensive attention from researchers within the field of intelligent dialogue systems. Developing dialogue agents that can persuade others to accept certain standpoints is essential to achieving truly intelligent and anthropomorphic dialogue systems. Benefiting from the substantial progress of Large Language Models (LLMs), dialogue agents have acquired an exceptional capability in context understanding and response generation. However, as a typical and complicated cognitive psychological system, persuasive dialogue agents also require knowledge from the domain of cognitive psychology to attain a level of human-like persuasion. Consequently, the cognitive strategy-enhanced persuasive dialogue agent (defined as CogAgent), which incorporates cognitive strategies to achieve persuasive targets through conversation, has become a predominant research paradigm. To depict the research trends of CogAgent, in this paper, we first present several fundamental cognitive psychology theories and give the formalized definition of three typical cognitive strategies, including the persuasion strategy, the topic path planning strategy, and the argument structure prediction strategy. Then we propose a new system architecture by incorporating the formalized definition to lay the foundation of CogAgent. Representative works are detailed and investigated according to the combined cognitive strategy, followed by the summary of authoritative benchmarks and evaluation metrics. Finally, we summarize our insights on open issues and future directions of CogAgent for upcoming researchers.
Introduction: Beyond the "Stochastic Parrot"
While Large Language Models (LLMs) have mastered the art of fluent conversation, they often suffer from the "Stochastic Parrot" problem—predicting the next token with high accuracy but lacking a true understanding of human cognitive states. In the realm of persuasion, mere linguistic fluency is insufficient. Effective persuasion requires a "Theory of Mind" (ToM), an understanding of the listener's psychological state, and the ability to plan a conversation several steps ahead to influence behavior or viewpoints.
This paper explores the emergence of CogAgents—dialogue systems that bridge the gap between linguistic capability and psychological strategy. By integrating cognitive psychology theories, these agents move from being reactive chatters to proactive influencers in high-stakes domains like mental health counseling, negotiation, and social good.
The CogAgent Framework
The core of the CogAgent paradigm is a modular architecture that "steers" LLM outputs using cognitive strategy mining and prediction. Unlike standard chatbots that respond turn-by-turn, CogAgents utilize a structured approach to dialogue modeling.
As shown in the architecture above, the system is divided into three primary phases:
- Cognitive Strategy Mining: Extracting patterns from existing human-to-human persuasive dialogues, including argument structures and topic transitions.
- Cognitive Strategy Prediction: Using the mined data to plan the next move in a dialogue, ensuring the agent remains goal-oriented.
- Application Scenarios: Deploying these strategies in specialized fields such as psychological counseling, debate, and negotiation.
Core Cognitive Strategies
The intelligence of a CogAgent is defined by three core strategies:
- Persuasion Strategy: Utilizing specific linguistic techniques derived from psychology, such as "logical appeal," "expression of empathy," or "task inquiry." These techniques are designed to lower the user's psychological defenses and build rapport.
- Topic Path Planning Strategy: A navigation tool that ensures the conversation flows logically toward a target. This often employs Reinforcement Learning (RL) or Monte Carlo Tree Search (MCTS) to optimize for long-term influence.
- Argument Structure Prediction Strategy: The ability to organize evidence—claims, premises, and rebuttals—into a logically sound case that can withstand user pushback.
Benchmarks and State-of-the-Art Performance
Research in CogAgents is supported by a diverse taxonomy of datasets across five major scenarios:
- Psychological Counseling: Datasets like ESConv and PsyQA focus on emotional support. Models such as MODERN and MultiESC lead this field by using knowledge-enriched encoders to perceive dynamic emotional changes.
- Negotiation and Social Good: CraigslistBargain and PersuasionForGood (charity donations) test an agent's ability to reach favorable outcomes.
- Recommendation: In movie recommendation tasks (REDIAL, INSPIRED), the LATTE framework has achieved superior results by pre-training modules as domain experts.
A significant trend in evaluation is the shift toward LLM-as-a-Judge. Since traditional metrics like BLEU (word overlap) cannot measure how "convincing" a response is, researchers increasingly use models like GPT-4 to score agents on strategy selection and emotional resonance.
Future Directions and Ethical Considerations
Despite rapid progress, several frontiers remain:
- Data Scarcity: There is a critical need for larger, richly annotated datasets that label specific cognitive strategies beyond simple intent.
- Multi-party Dynamics: Most research focuses on 1-on-1 interactions, but real-world persuasion often involves groups, requiring agents to manage multiple "Theories of Mind."
- Interpretability: Combining Knowledge Graphs with LLMs is a promising path to make the agent's persuasive logic more transparent.
- Ethics: The ability to influence human behavior raises significant concerns regarding manipulation. Future research must prioritize robust safety frameworks and ethical guidelines to ensure CogAgents are used for social good rather than exploitation.
Summary updated