ARES: A Multi-Agent Collaboration-Driven Automatic Heuristic Algorithm Evolution System

ARES (AI Research Ensemble System) is a system that realizes automatic algorithm evolution and discovery using large language models (LLMs) through the collaboration of three roles: Theorist, Critic, and Experimenter. It aims to provide intelligent solutions for complex optimization problems and replace traditional algorithm design methods that rely on human experts' experience.

Heuristic algorithms are crucial for complex optimization problems (such as the Traveling Salesman Problem and scheduling optimization). However, designing efficient algorithms relies on deep domain knowledge and experience, and different problems require different strategies, making manual design costly. In recent years, with the improvement of LLM capabilities, researchers have explored using them to automatically generate optimization code. ARES simulates the collaboration of a research team to achieve algorithm evolution.

The innovation of ARES lies in its three-role architecture:

• Theorist: Generates and maintains a strategy table, proposes improvement hypotheses and experimental plans;
• Critic: Validates hypotheses through structural ablation experiments and parameter scanning, and quickly feeds back excellent individuals;
• Experimenter: Converts strategies into executable code. The evolution cycle includes initialization (loading configurations, generating initial strategy tables and populations), main cycle (updating strategies → validation → integration → next generation), elite retention, aggressive exploration (safe high-risk exploration), and meta-reflection (distilling experience).

Evaluation Mechanism: Candidate programs are executed in an isolated temporary directory, and the last line of stdout is scanned in reverse to get the floating-point number as the score; Fault tolerance: Timeouts, exceptions, etc., are marked as failures and eliminated with an infinite value. Configuration System: Parameters such as the number of parallel processes and timeouts are controlled via cfg/config.yaml, supporting quick adaptation to new problems. Output Structure: A timestamped working directory stores outputs from each role, weight history, etc., ensuring traceability.

Application Scenarios of ARES:

• Algorithm Researchers: Assist in exploring new algorithms and validating hypotheses;
• Engineers: Automatically discover efficient heuristic algorithms for business problems;
• Educators: Demonstrate the application of multi-agent collaboration and evolutionary computing. It represents a new model of AI-assisted scientific research, combining LLMs with scientific methods, and can be extended to fields requiring creative thinking and rigorous validation.

Limitations: Relies on LLM code generation capabilities (complex logic is prone to errors), high computational resource requirements, only supports single-objective optimization, and the strategy table design needs optimization. Future Directions: Introduce more validation mechanisms to improve code reliability, explore efficient evaluation strategies to reduce overhead, extend to multi-objective/constraint optimization, and combine with AutoML to form a complete solution.

ARES demonstrates the potential of LLMs in the field of automatic algorithm discovery. By building a self-evolving system through multi-role collaboration, it is an exploration of a new paradigm for AI-assisted scientific research. With the improvement of LLM capabilities and system optimization, it is expected to solve more complex optimization challenges and accelerate the process of scientific discovery.