ARES: Autonomous rule-discovery evolutionary system for swarm intelligence driven by the synergy of knowledge and experience

To address the limitations of large-language-model-based automated algorithm design, particularly the lack of long-term reflection and inefficient blind search, this study proposes ARES, a knowledge-driven framework that combines the semantic reasoning of large language models (LLMs) with the iterative search of evolutionary computation. ARES maintains a dynamic knowledge base that integrates theoretical priors with empirical evolutionary experience, enabling reusable knowledge to guide later generations. It further uses a hierarchical strategy–tactics–coding planning schema and a closed-loop evaluation mechanism to generate executable control code. By accumulating trial-and-error experience as structured knowledge, ARES moves the search from unguided mutation toward knowledge-guided evolution. Experiments on four swarm tasks show that ARES improves final performance and behavioral stability over the compared baselines while producing interpretable control strategies.