Learning to Align Decentralized Agents with Global Goals via Co-Evolution of Code and Language
Preparing for submission, 2025
The combination of Large Language Models (LLMs) and Evolutionary Algorithms (EAs) offers a new paradigm for solving complex, real-world problems. We introduce Individual-Global Alignment via Evolution of Heuristics (IGA-EH), a framework that leverages LLMs as intelligent variation and interpretation engines within an evolutionary loop to align decentralized agent behavior with global system objectives. IGA-EH simultaneously evolves executable decision-making heuristics and natural language nudges, enabling adaptive mechanism design that integrates behavioral and programmatic reasoning. Applied to agricultural landscapes, the framework discovers heuristics that approximate optimal ecological connectivity while generating persuasive messages that steer diverse simulated agents toward collective outcomes. Crucially, IGA-EH extends beyond prior LLM-EA work focused on benchmark tasks by addressing real-world, mixed-integer optimization problems and co-evolving both behavioral rules and communication strategies. This approach establishes a generalizable method for influencing agent behavior in complex systems, with applications in environmental planning, resource governance, and aligned AI decision-making.