AgentOPSD: Recursive Self-Distillation for Agentic Reinforcement Learning

AgentOPSD introduces a critic-free, recursive method for turn-level credit assignment in agentic reinforcement learning, utilizing Bayesian belief states to address long-horizon tasks. Simultaneously, research into visual continuous control highlights the challenges of sample-efficient policy learning from pixels. While dynamics-based representation learning uses self-prediction or observation prediction, current methods struggle with limited data. The proposed observation-grounded self-predictive reinforcement learning aims to improve these representations by combining predictive objectives to enhance overall performance.