Imfo Hub Distinguished Lecture Series - CALM: Co-evolution of Algorithms and Large Language Models for Automatic Heuristic Design

This talk presents a novel methodology for advancing solutions to a broad range of challenging optimization problems. Traditionally, solving complex optimization tasks has depended heavily on expert-designed heuristics, which are often developed through costly and time-consuming trial-and-error processes. Recent progress shows that large language models (LLMs), when embedded within well-designed evolutionary search frameworks, can automatically discover high-performing heuristics with far less effort. However, existing approaches largely depend on verbal guidance, i.e., manipulating prompts to steer the evolution of heuristics, without adapting the LLM itself. In this talk, we propose a hybrid framework that integrates both verbal and numerical guidance. Specifically, we fine-tune the LLM through reinforcement learning (RL), leveraging feedback on the quality of generated heuristics to enable the LLM to co-evolve alongside the search process. This joint optimization yields superior performance, consistently outperforming state-of-the-art baselines across various optimization tasks. At the end of the talk, we will discuss several potential applications of CALM in solving real-world engineering problems, including, for example, network resource allocation and network security.