EvoSR-LLM: Evolutionary Symbolic Regression Guided by Large Language Models
提出EvoSR-LLM方法，结合进化算法与大语言模型，通过自适应提示机制生成更有效的数学方程，在24个科学基准问题上优于现有方法。
Mathematical equations describe fundamental laws across various disciplines, yet discovering concise and effective mathematical expressions from data remains a challenging task. Traditional symbolic regression methods often make limited use of the domain-specific prior knowledge, while large language model (LLM)-based symbolic regression approaches can more effectively utilize it. However, the existing LLM-based symbolic regression methods require substantial computational resources to generate equations while still suffering from low efficiency in producing effective ones. To address this issue, we propose EvoSR-LLM, a prompt-guided evolutionary search algorithm for equation discovery. EvoSR-LLM represents each equation as an individual composed of knowledge, insights, and code, integrating an evolutionary algorithm for iterative refinement. By leveraging an adaptive prompt mechanism, EvoSR-LLM constructs diverse prompts to guide the LLM in generating more effective equations. Experimental results demonstrate that EvoSR-LLM outperforms state-of-the-art symbolic regression baselines on 24 scientific benchmark problems, discovering equations with higher accuracy, stronger generalization, and lower complexity.