Task-oriented dialogue systems have recently begun to support dynamic user goals, allowing agents to react to evolving preferences and sentiment shifts during a conversation. However, even state-of-the-art dynamic goal-driven agents (e.g., DGDVA) remain largely reactive: when a user updates their goal to something infeasible given domain constraints (such as requesting flagship specifications at a budget price), these systems can detect the mismatch but cannot actively guide the user toward a realistic alternative. In this paper, we introduce PDG-DVA (Persuasive Dynamic Goal-Driven Virtual Agent), a dialogue framework that couples dynamic goal adaptation with computational persuasion to salvage conversations that would otherwise fail. We formalize the problem as an augmented POMDP whose state includes a persuasion state (activation flag, chosen strategy, rejection count, and estimated user flexibility), and we extend the action space with persuasion-specific actions such as compromise suggestion and value reframing. A tripartite reward—covering task completion, sentiment alignment, and successful persuasive redirection—encourages the agent to solve feasible tasks normally while invoking persuasion only when infeasibility is detected. To support training and evaluation, we release DevVA-P, a 250-dialogue extension of the DevVA dataset annotated with infeasible goal turns, applicable persuasion strategies, and acceptance outcomes. Experiments on DevVA-P show that PDG-DVA substantially improves persuasion acceptance rate and average return over strong baselines, including the original DGDVA, a random-persuasion variant, and a rule-based persuader, while human evaluation confirms that users perceive the system as more helpful and intelligent without sacrificing naturalness.
Task-oriented dialogue systems have recently begun to support dynamic user goals, allowing agents to react to evolving preferences and sentiment shifts during a conversation. However, even state-of-the-art dynamic goal-driven agents (e.g., DGDVA) remain largely reactive: when a user updates their goal to something infeasible given domain constraints (such as requesting flagship specifications at a budget price), these systems can detect the mismatch but cannot actively guide the user toward a realistic alternative. In this paper, we introduce PDG-DVA (Persuasive Dynamic Goal-Driven Virtual Agent), a dialogue framework that couples dynamic goal adaptation with computational persuasion to salvage conversations that would otherwise fail. We formalize the problem as an augmented POMDP whose state includes a persuasion state (activation flag, chosen strategy, rejection count, and estimated user flexibility), and we extend the action space with persuasion-specific actions such as compromise suggestion and value reframing. A tripartite reward—covering task completion, sentiment alignment, and successful persuasive redirection—encourages the agent to solve feasible tasks normally while invoking persuasion only when infeasibility is detected. To support training and evaluation, we release DevVA-P, a 250-dialogue extension of the DevVA dataset annotated with infeasible goal turns, applicable persuasion strategies, and acceptance outcomes. Experiments on DevVA-P show that PDG-DVA substantially improves persuasion acceptance rate and average return over strong baselines, including the original DGDVA, a random-persuasion variant, and a rule-based persuader, while human evaluation confirms that users perceive the system as more helpful and intelligent without sacrificing naturalness.
