Abstract
Methods for generating proactive conversational follow-up questions, or nudges, may be limited by a reliance on heuristic rules or subjective evaluations, which can correlate poorly with actual user engagement. The described technology may use a two component system. A predictive model, trained on historical user interaction data, can be configured to generate a quantitative score representing the likelihood of user acceptance for a given nudge. This score can serve as a reward signal for a generative response planner operating within a reinforcement learning framework. The planner can iteratively refine system instructions for a separate, downstream response-generation language model. This data-driven feedback loop can facilitate a scalable and automated process for creating conversational nudges that may have a higher probability of user acceptance, which may improve multi-turn dialogue.
Creative Commons License

This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Bailey, Eric; Jun, Changgyoung; and Macrì, Tommaso, "System for Optimizing Conversational Prompts via Reinforcement Learning Guided by an Interaction Prediction Model", Technical Disclosure Commons, (November 20, 2026)
https://www.tdcommons.org/dpubs_series/13181