↓Skip to main content
  1. Projects/

QD-LEAR - Quality-Diversity in LLM-Evolved Agents

·2 mins
Research AI Agents Evolutionary Algorithms Quality-Diversity
Table of Contents

Quality-Diversity Tradeoffs in LLM-Evolved Agents #

QD-LEAR extends the LEAR (LLM-Driven Evolution of Agent-Based Rules) research to explore quality-diversity tradeoffs in evolved agent rule representations. This work investigates how to enhance behavioral diversity in multi-agent systems while maintaining performance quality.

Research Background #

Building on LEAR’s success in using LLMs to evolve agent behaviors, QD-LEAR addresses a critical challenge: evolved agents often converge to narrow behavioral optima, limiting exploration of the solution space. This research applies quality-diversity algorithms from evolutionary computation to LLM-driven agent evolution.

Key Contributions #

Methodological Innovation #

  • Integrated Map-Elites algorithm with LLM-driven evolution
  • Developed behavioral descriptors for agent rule diversity
  • Created multi-objective evaluation balancing quality and diversity
  • Systematic exploration of QD algorithms in agent-based modeling

Technical Advances #

  • Novel behavioral space characterization for agent rules
  • Adaptive mutation strategies guided by diversity metrics
  • Archive maintenance for diverse high-performing agents
  • Comparative analysis of QD approaches in LLM evolution

Research Questions #

  1. How can we maintain behavioral diversity while evolving agents with LLMs?
  2. What behavioral descriptors best capture agent rule diversity?
  3. How do different QD algorithms compare in LLM-driven evolution?
  4. When does diversity matter for agent system performance?

Experimental Design #

Test Beds #

  • NetLogo multi-agent simulations
  • Various scenario types (resource collection, predator-prey, cooperation)
  • Baseline comparison with standard LEAR

Metrics #

  • Behavioral diversity measurement
  • Archive coverage in behavioral space
  • Performance across diverse scenarios
  • Convergence analysis

Technology Stack #

  • Python for experimental framework
  • LLM APIs (Claude, Grok, DeepSeek) for rule generation
  • NetLogo for agent simulation
  • Custom QD algorithm implementations
  • Jupyter notebooks for analysis

Publication #

Accepted at ALife 2025 - Conference on Artificial Life

  • Peer-reviewed publication in artificial life research
  • Builds on GECCO ‘25 LEAR publication
  • Contributed to understanding of diversity in AI evolution

Impact #

This research advances the state of the art in LLM-driven evolution by demonstrating how quality-diversity approaches can enhance agent exploration and prevent premature convergence. The work has implications for evolving more robust and adaptable multi-agent systems.