Nudging towards Sustainability: Persona-Driven Reinforcement Learning for Empowering Informative Conversation

ACL ARR 2025 July Submission1360 Authors

Abstract: This research introduces a novel dataset, EcoNudge, and a persona-aware RL framework designed to empower conversational agents in promoting sustainability through subtle, personalized guidance. Our core contributions are the EcoNudge dataset, tailored for the sustainability domain with specific personas and interaction design, and an RL methodology that adapts established techniques for this nuanced task. The novelty lies in this specific application to sustainability dialogues, the characteristics of the dataset itself, and the insights which demonstrate that RL-enhanced smaller language models can achieve strong performance, specifically and on key task-specific guidance metrics (Guidance Efficacy, Support Consistency), compared to larger, general-purpose prompted LLMs. This underscores the value of targeted RL fine-tuning for specialized applications like nudging. This work addresses a societal need for more effective and personalized communication tools to encourage sustainable practices. Future work includes exploring more diverse cultural personas, and undertaking longitudinal studies on real-world impact.

Paper Type: Long

Research Area: Dialogue and Interactive Systems

Research Area Keywords: NLP datasets, Dialogue Systems, Applications, Human-in-the-loop

Contribution Types: Data resources

Languages Studied: English

Submission Number: 1360