Don't Wait to Reply: Towards Responsive yet Thoughtful Dialogue through Proactive Thinking

ACL ARR 2026 January Submission5542 Authors

Keywords: proactive thinking, dialogue system, large language model

Abstract: While chain-of-thought reasoning has significantly advanced the reasoning capabilities of Large Language Models (LLMs), its sequential reactive thinking nature introduces substantial latency, which can degrade the responsiveness and fluidity of conversational systems. In this position paper, we propose proactive thinking—a paradigm inspired by human conversational dynamics in which LLMs perform reasoning during natural dialogue intervals or while the user is speaking. This approach enables the model to pre-plan elements of its response before its turn begins. We argue for its feasibility from a predictive branching perspective. As preliminary validation, we implement a prototype proactive-thinking system using prompting, demonstrating that LLMs can effectively plan upcoming responses in task-oriented dialogues, even without knowing the user's next utterance. Our work advocates for a shift toward more intelligent, real-time interaction models in future conversational AI.

Paper Type: Long

Research Area: Dialogue and Interactive Systems

Research Area Keywords: task-oriented, conversational modeling, spoken dialogue systems

Contribution Types: Position papers

Languages Studied: English

Submission Number: 5542