The paper presents AnyMDP, a framework for procedurally generating diverse tasks to enhance In-Context Reinforcement Learning (ICRL) scalability, and explores the trade-off between generalization and adaptation efficiency.