A Review on Systematically Enhancing the Application Capabilities of Large Language Models

The advent of Large Language Models (LLMs) has instigated a reconsideration of the scientific research paradigm. These potent LLMs, exemplified by those released by leading companies and research institutions such as OpenAI, have exhibited remarkable capabilities. They encapsulate an extensive array of world knowledge and exhibit robust language generation, context learning, reasoning, few-shot or zero-shot abilities, and generalization traits. They are anticipated to tackle intricate tasks and problems. To augment the capabilities of these LLMs, numerous methods have been suggested, including pre-training and fine-tuning, prompt engineering, the use of external tools, and multimodality. These strategies have substantially enhanced the performance of LLMs. At present, these models have found applications in various domains, including prediction and embodied intelligence. Nonetheless, the potential of LLMs is yet to be fully realized, and their evolution is confronted with several challenges and issues. This paper delineates the strategies to enhance the applicability of LLMs, reviews the methods and research advancements, and ultimately proposes a comprehensive system-level application framework to facilitate the superior application of LLMs in diverse scenarios.