Evolving Personalities in Chaos: An LLM-Augmented Framework for Character Discovery in the Iterated Prisoners Dilemma under Environmental Stress

TL;DR

This work tackles two core IPD limitations—environmental realism and interpretability—by introducing stochastic God Mode perturbations during evolution and an LLM-based behavioral profiler that converts evolved genotypes into narrative character profiles. It evolves 18-bit LUT-based strategies with standard GA operators under five perturbations, then uses four behavioral tests to extract metrics for GPT-5.1 to generate interpretable archetypes (e.g., The Paranoid Pacifist). Results show chaotic environments promote resilient, forgiving strategies and demonstrate that LLM-generated profiles align with observed behaviors, providing a bridge between evolutionary computation and explainable AI. The approach offers a reproducible template for automated, human-understandable agent characterization in multi-agent systems.

Abstract

Standard simulations of the Iterated Prisoners Dilemma (IPD) operate in deterministic, noise-free environments, producing strategies that may be theoretically optimal but fragile when confronted with real-world uncertainty. This paper addresses two critical gaps in evolutionary game theory research: (1) the absence of realistic environmental stressors during strategy evolution, and (2) the Interpretability Gap, where evolved genetic strategies remain opaque binary sequences devoid of semantic meaning. We introduce a novel framework combining stochastic environmental perturbations (God Mode) with Large Language Model (LLM)-based behavioral profiling to transform evolved genotypes into interpretable character archetypes. Our experiments demonstrate that strategies evolved under chaotic conditions exhibit superior resilience and present distinct behavioral phenotypes, ranging from Ruthless Capitalists to Diplomatic Enforcers. These phenotypes are readily classified by LLMs but remain nearly impossible to interpret through manual genome inspection alone. This work bridges evolutionary computation with explainable AI and provides a template for automated agent characterization in multi-agent systems.

Figures (4)

• Figure 1: Overview of the Evolutionary IPD framework showing the three-tier architecture: Streamlit UI for visualization, Simulation Engine with stochastic stressor integration, and Analysis Engine for behavioral profiling and LLM character generation.
• Figure 2: Comparison of fitness trajectories over 100 generations. The sterile environment (blue) achieves higher absolute fitness through rapid optimization without penalty, while the chaotic environment (red) maintains higher variance throughout evolution. The stochastic stressors impose direct fitness penalties, resulting in lower average fitness but selecting for strategies resilient to perturbation.
• Figure 3: Complete profile of "The Paranoid Pacifist" champion agent. (a) Character card in collectible card style, generated using GPT-5.2 with the LLM-synthesized character attributes (name, motto, alignment, description). (b) Genotype heatmap visualization showing the 18-gene response pattern, where green indicates cooperation and red indicates defection.
• Figure 4: Population-level cooperation rate over 100 generations. Both environments begin at approximately 50% cooperation. The sterile environment (blue) maintains relatively stable cooperation near the 50% baseline. The chaotic environment (red) shows a pronounced initial decline (Gen 0-35), followed by a partial recovery phase (Gen 35-70), before declining again---suggesting sustained environmental stress continues to select for defection-dominant strategies.