A multi-agent framework powered by large language models for automatic heuristic design

Heuristic design for combinatorial optimization problems has long relied on human expertise and repeated trial-and-error. Recent large language model (LLM)-based approaches to Automatic Heuristic Design have made notable progress. Most existing methods still remain within performance-driven generate-and-evaluate loops, where mechanism-level validation and cross-generational knowledge accumulation are only weakly developed. To address this limitation, we propose ARES (AI Research Ensemble System). ARES organizes Automatic Heuristic Design as a validation-driven and knowledge-accumulative multi-module collaborative workflow. It consists of a Theorist for mechanism-level analysis, an Experimenter for candidate generation, and a Critic module for structural ablation and parameter-scanning-based validation. It also introduces a Strategy Table to support cross-generational inheritance of mechanism-level design experience. Experiments on six representative combinatorial optimization problems show that ARES attains the best mean performance among the compared methods under the reported settings. Across these six structurally diverse problems, ARES also shows consistent improvements within the evaluated benchmark set.