deployment-time-learning

This paper introduces CASCADE, a framework for deployment-time learning that allows Large Language Models to adapt continuously through episodic memory and contextual bandit optimization without modifying model parameters.