CASCADE: Case-Based Continual Adaptation for Large Language Models During Deployment

TL;DR

CASCADE introduces a framework for large language models to learn and adapt during deployment through episodic memory, improving performance across diverse tasks without retraining.

Contribution

The paper formalizes deployment-time learning for LLMs and proposes CASCADE, enabling continual adaptation via episodic memory and contextual bandit formulation.

Findings

CASCADE improves success rate by 20.9% over zero-shot prompting.

It outperforms gradient-based and memory-based baselines.

Demonstrates effectiveness across 16 diverse tasks.

Abstract

Large language models (LLMs) have become a central foundation of modern artificial intelligence, yet their lifecycle remains constrained by a rigid separation between training and deployment, after which learning effectively ceases. This limitation contrasts with natural intelligence, which continually adapts through interaction with its environment. In this paper, we formalise deployment-time learning (DTL) as the third stage in the LLM lifecycle that enables LLM agents to improve from experience during deployment without modifying model parameters. We present CASCADE (CASe-based Continual Adaptation during DEployment), a general and principled framework that equips LLM agents with an explicit, evolving episodic memory. CASCADE formulates experience reuse as a contextual bandit problem, enabling principled exploration-exploitation trade-offs and establishing no-regret guarantees over long-term interactions. This design allows agents to accumulate, select, and refine task-relevant cases, transforming past experience into actionable knowledge. Across 16 diverse tasks spanning medical diagnosis, legal analysis, code generation, web search, tool use, and embodied interaction, CASCADE improves macro-averaged success rate by 20.9% over zero-shot prompting while consistently outperforming gradient-based and memory-based baselines. By reframing deployment as an adaptive learning process, this work establishes a foundation for continually improving AI systems.