Some Optimization Problems in Large Language Models

Large language model (LLM) is a research hotspot in the field of artificial intelligence (AI). Currently, various LLMs have been designed and have demonstrated outstanding language understanding and generation capabilities in various natural language processing tasks. With the development of LLM research, a spectrum of optimization problems have emerged, posing challenges to the pursuit of further performance gains of LLMs. To enlighten future research on the optimization problems in LLMs, this paper summarizes some significant optimization problems in LLMs and proposes a classification method based on the model’s scope to categorize them. Specifically, this paper focuses on three types of cutting-edge issues namely model merging, prompt engineering, and jailbreaking attack. First, the model merging refers to the optimization problem of the integration of different pre-trained models. Second, the prompt engineering refers to the optimization problem that involves designing effective input prompts to guide LLMs to generate desired responses. Third, the jailbreaking attack refers to the optimization problem that indicates the process of crafting adversarial inputs that can circumvent the safety constraints or ethical guidelines programmed into LLMs. Also, we deeply analyze how evolutionary computation (EC) can empower LLMs with customized optimization solutions through their autonomous learning and efficient search capabilities, significantly improving the automation efficiency. Based on the above analyses, we finally discuss some potential future problem directions of LLM, aiming at promoting the deep integration and innovative development of LLMs in the field of automated optimization, leading to the next generation of AI.