Jie Shi

Standard single-turn RL produces socially myopic agents, but dynamically scheduling rewards from early relationship-building to mid-dialogue task execution drives a 9.2 percentage point gain in multi-turn goal achievement.