Challenges and Strategies in the Development of Large Models

In recent years, the rapid advancement of Large Models (LMs), including large language models, visual foundation models, and multimodal LMs, has significantly transformed various fields. These models are evolving at a very fast pace, and their development has benefited numerous industries considerably. However, inherent architectural limitations in LMs, such as hallucinations and challenges in error localization, restrict their potential. Addressing these issues effectively is crucial for their continued progress. This work provides an overview of the evolution of LMs and highlights key challenges they face, including high data and energy demands, catastrophic forgetting, limited reasoning capabilities, and fault localization. Strategies to mitigate these challenges are proposed, followed by a discussion on applying LMs in smart industrial production. Harnessing the strengths of LMs is expected to unlock new opportunities and drive innovation across various industries.