CASCADE: A Framework for Continual LLM Adaptation During Deployment

A new paper on arXiv introduces CASCADE (CASe-based Continual Adaptation during Deployment), a framework for deployment-time learning (DTL) in large language models (LLMs). The authors formalize DTL as a third stage in the LLM lifecycle, enabling agents to improve from experience without modifying model parameters. CASCADE equips LLM agents with an explicit, evolving episodic memory and formulates experience reuse as a contextual bandit problem, allowing principled exploration-exploitation trade-offs and no-regret guarantees over long-term interactions. The approach addresses the limitation of current LLMs, which cease learning after deployment, contrasting with natural intelligence's continual adaptation. The paper is available on arXiv with ID 2605.06702.