Evolutionary ecology of words

TL;DR

Problem: study word-level evolutionary dynamics using LLMs to generate, judge, and mutate words. Approach: a minimal agent-based framework with $N$ agents on a $W \times W$ toroidal grid where each carries a word drawn from an initial list of size $A$; interactions are resolved by LLM judgments and mutations occur with probability $p_m$ to one of $B$ variants, all guided by prompts. Key findings: starting from a population of well-known species, diverse ecological lineages—terrestrial, marine, extinct—emerge and exhibit punctuated equilibria; a large-population long-term run yields thousands of unique species, e.g., $3{,}704$, with ongoing coexistence and dynamic turnover, driven by LLM-generated relationships. Significance: demonstrates that LLMs can serve as engines for open-ended evolution in artificial life, generating novel, linguistically mediated ecological dynamics with potential applications in complex-systems research.

Abstract

We propose a model for the evolutionary ecology of words as one attempt to extend evolutionary game theory and agent-based models by utilizing the rich linguistic expressions of Large Language Models (LLMs). Our model enables the emergence and evolution of diverse and infinite options for interactions among agents. Within the population, each agent possesses a short word (or phrase) generated by an LLM and moves within a spatial environment. When agents become adjacent, the outcome of their interaction is determined by the LLM based on the relationship between their words, with the loser's word being replaced by the winner's. Word mutations, also based on LLM outputs, may occur. We conducted preliminary experiments assuming that ``strong animal species" would survive. The results showed that from an initial population consisting of well-known species, many species emerged both gradually and in a punctuated equilibrium manner. Each trial demonstrated the unique evolution of diverse populations, with one type of large species becoming dominant, such as terrestrial animals, marine life, or extinct species, which were ecologically specialized and adapted ones across diverse extreme habitats. We also conducted a long-term experiment with a large population, demonstrating the emergence and coexistence of diverse species.

Figures (5)

• Figure 1: Overview of the model. (a) Individual (word) placement in 2D space, (b) word list generation for the initial population, (c) movement and LLM-based competition, (d) payoff matrix reflecting the competitive relationship, and (e) mutation example showing possible word variations.
• Figure 2: Prompts for "strong animal species" survives. The base prompt (a), defines the basic behavior of the LLM, is prepended to each prompt for specific process: (b) initial word list generation, (c) competition judgment, and (d) mutation generation.
• Figure 3: Transition of populations in the semantic space of words over 10 trials. Each dot represents the average word vector for each trial, plotted every 20 steps, with words vectorized in 2D using SentenceTransformer and UMAP. The 10 most frequent words in each trial (gray) and the most frequent word in the final step (bold black) are shown.
• Figure 4: Spatial distribution of individuals in the final step of Trial 2 (top left), word frequency in the final step (top right), and transition of frequencies for the top 10 species through the trial (bottom).
• Figure 5: Transition of populations in the semantic space of words in a long-term trial. Each dot represents the average word vector plotted every 20 steps with color gradient (blue to red) indicating temporal progression. In order to illustrate the species diversity of the population, the word distributions are shown every 100 steps with the same color gradient indicating temporal progression. The most frequent word in the final step (bold black), 50 most frequent species (gray), and the dominant species every 100 steps (in blue) are shown.