CDEoH: Category-Driven Automatic Algorithm Design With Large Language Models

A research team led by Yu-Nian Wang has introduced CDEoH (Category-Driven Automatic Algorithm Design with Large Language Models), a novel framework that tackles a core weakness in using LLMs to generate new algorithms. Current LLM-based heuristic search methods often suffer from unstable evolutionary processes and get stuck in suboptimal solutions (premature convergence). While existing fixes focus on prompt engineering or evolving code and reasoning together, CDEoH identifies a missing piece: the lack of algorithmic category diversity. It argues that maintaining a population of solutions from varied algorithmic paradigms is critical for stable, long-term evolution.

The CDEoH framework explicitly models different algorithm categories and implements a population management strategy that balances two objectives: raw performance and category diversity. This dual focus allows the system to explore multiple promising evolutionary directions in parallel, rather than having the entire population converge prematurely on a single approach. The researchers validated CDEoH through extensive experiments on representative combinatorial optimization problems at multiple scales. The results demonstrated that the method effectively mitigates single-direction convergence, significantly enhances evolutionary stability, and achieves consistently superior average performance across different tasks and problem sizes compared to previous approaches.