look-ahead-simulations

This paper introduces Preference Tree Optimization (PTO), a framework that generates preference data via look-ahead simulations to iteratively improve goal-oriented dialogue agents, with experiments showing gains in Motivational Interviewing settings.