From Static Context to Calibrated Interactive RL: Mitigating Distribution Shift in Multi-turn Dialogue with Aligned Simulator

In three linesTheoretical and empirical work on training LLM-based dialogue agents. Identifies context distribution shift as fundamental limitation of Static Context RL and Interactive RL. Proposes Calibrated Interactive RL combining interactive RL with simulator alignment to reduce sim-to-real gap and improve multi-turn dialogue quality.

Summary generated by Claude — human-verified