CUP Framework: Uncertainty-Aware Planning for Goal-Oriented Conversations

A new framework, Conversation Uncertainty-aware Planning (CUP), has been proposed to improve goal-oriented conversational systems by integrating language models with structured planning. The research, detailed in arXiv paper 2604.03924, addresses the challenge of sequential decision-making under uncertainty about user intent. Existing methods either rely on predefined schemas (structured) or lack long-horizon decision-making (LLM-based), leading to poor coordination between information acquisition and target commitment. CUP formulates goal-oriented conversation as an uncertainty-aware sequential decision problem, using uncertainty as a guiding signal for multi-turn decision making. The framework has a language model propose feasible actions, and a planner evaluates them, balancing information gathering and commitment. This approach aims to enhance coordination in multi-turn conversations, potentially improving performance in tasks like negotiation, recommendation, and customer support. The paper is available on arXiv and was last updated with a replace-cross announcement type.