IB-RL: Isolated Bilateral Reinforcement Learning for Strategic Dialogue Agents

IB-RL addresses the static-counterpart mismatch in strategic dialogue training by preventing target agents from exploiting fixed simulator regularities. The method employs bilateral reinforcement learning to improve policy generalization across diverse, adaptive opponents rather than optimizing against a stationary environment.