LEAR - LLM-Driven Evolution of Agent-Based Rules

🧬 Evolutionary Agent-Based Modeling with LLMs

LEAR (LLM-Driven Evolution of Agent-Based Rules) is a cutting-edge research project that explores using Large Language Models within Agent-Based Modeling environments to iteratively enhance agent movement and functionality through automated code generation. This work was accepted to GECCO ‘25.

Research Overview

Core Innovation

This project introduces a novel methodology where evolution operates at a higher abstraction level by mutating pseudocode representations of agent behaviors, subsequently converting them into executable code through LLM-mediated steps.

Key Contributions

• Semantic Evolution: Leverages LLMs to introduce semantically meaningful variations during evolution
• Abstraction-Level Mutation: Operates on pseudocode for more innovative solutions
• Systematic Comparison: Evaluates different prompting strategies for code generation
• Multi-Agent Benchmarks: Provides comprehensive evaluation frameworks

Technical Approach

Architecture Components

LLM-Driven Evolution

# Evolution Pipeline
1. Generate initial agent behaviors
2. Evaluate performance in NetLogo environment
3. Use LLM to propose semantic mutations
4. Convert pseudocode to executable code
5. Iterate through generations

Prompting Strategies

• Direct Code Generation: LLM generates executable code directly
• Pseudocode Intermediate: Evolution on abstract representations
• Guided Mutation: Semantically meaningful variation operators
• Fitness-Informed: Performance metrics guide evolution

Multi-Agent Domains

• Movement optimization in complex environments
• Collective behavior emergence
• Resource allocation strategies
• Swarm intelligence patterns

Implementation Stack

Research Impact

Academic Contributions

• Publication: Accepted at Genetic and Evolutionary Computation Conference (GECCO ‘25)
• Novel Methodology: First systematic exploration of LLM-driven evolution in ABM
• Open Source: Complete implementation and benchmarks available

Practical Applications

• Automated agent behavior design
• Complex system optimization
• Emergent behavior discovery
• Code synthesis for simulations

Experimental Results

Performance Metrics

• Demonstrates superior solution quality compared to traditional GP
• Achieves faster convergence through semantic mutations
• Produces more interpretable agent behaviors
• Scales effectively to complex multi-agent scenarios

Key Findings

• LLMs excel at introducing meaningful behavioral variations
• Pseudocode evolution enables discovery of novel strategies
• Natural language training data enhances solution creativity
• Systematic prompt engineering significantly impacts performance

Citation

@inproceedings{LEAR_GURKAN,
  author = {Gurkan, Can and Jwalapuram, Narasimha Karthik and 
            Wang, Kevin and Danda, Rudy and Rasmussen, Leif and 
            Chen, John and Wilensky, Uri},
  title = {LEAR: LLM-Driven Evolution of Agent-Based Rules},
  year = {2025},
  booktitle = {Proceedings of GECCO '25},
  publisher = {ACM},
  doi = {10.1145/3712255.3734368}
}

Repository