Large Language Models Are Semi-Parametric Reinforcement Learning Agents

As declared by Seifert et al. [1997], the episodic memory of the experiences from past episodes plays a crucial role in the complex decision-making processes of human [Suddendorf and Corballis, 2007]. By recollecting the experiences from past episodes, the human can learn from success to repeat it and learn from failure to avoid it.