ReGAP: Evidence-Aligned Conversational Question Generation via Reasoning-Guided Action Planning

Abstract

Conversational question generation (CQG) has shown promise for supporting learners in text-based dialogue. Existing methods mainly optimize question relevance, fluency, and answer consistency, but pay less attention to whether follow-up questions help learners revise incomplete or selective use of evidence. In practice, learners may rely on only a subset of task-relevant evidence, leading to incomplete or overgeneralized interpretations of the source content. To address this problem, we propose Reasoning-Guided Action Planning (ReGAP), a framework for evidence-aligned CQG. ReGAP formulates follow-up question generation as a sequential intervention planning problem. It first estimates the learner's evidence-utilization state from the dialogue history, then constructs structured intervention actions by pairing target evidence with reasoning operations. Based on this state-action representation, ReGAP uses Monte Carlo Tree Search to compare candidate intervention strategies over future dialogue trajectories, and further incorporates experience priors to improve planning efficiency and stability. Experiments on multiple dialogue datasets show that ReGAP consistently improves evidence-aligned reasoning over strong baselines. Expert annotation and real-user studies further demonstrate that ReGAP generates questions that better guide learners toward task-relevant evidence while maintaining conversational fluency and positive user experience. These results suggest that planning-based conversational intervention is a promising direction for evidence-aligned CQG.