Industrial LLMs Need Continual Learning Ecosystems for Sustainability

Large Language Models (LLMs) deployed in industrial settings require continuous updates to adapt to changing requirements and environments, rather than being retrained from scratch. This paper introduces the concept of Industrial Continual Learning (ICL) for LLMs, viewing it as a dynamic, versioned ecosystem where model updates propagate hierarchically. The authors highlight critical challenges such as the degradation of model plasticity over time, the disruption of capability inheritance during foundation model upgrades, and the constraints imposed by long-term deployment requirements. To address these issues, the survey outlines five key lifecycle design principles. These include strategies for preserving model plasticity, managing upgrades as a form of capability transfer, ensuring trustworthy continual reinforcement learning, developing self-optimizing training recipes, and establishing accountability for long-term iteration. The paper also assesses the current maturity of these principles, identifies gaps in real-world deployment, and proposes a practical blueprint for ICL implementation, aiming to bridge the gap between industrial needs and academic research.