Behavioral Inheritance and Evolution in LLM-Controlled Agent Populations

Abstract

Genomes in artificial evolutionary systems have taken many forms, among them bit strings, program instructions, and network weights, none of which makes the expected behavior apparent to a human being. Can a genome state its behavior directly, in ordinary human language, instead? We ask whether a corpus of natural-language behavioral instructions, carried by an agent whose actions are chosen by a large language model, supports faithful inheritance, environment-driven behavioral change, heritable mutation, and selection among competing variants of a gene. A positive answer gives artificial life a new kind of software organism, one whose genes are sentences, so that inheritance, expression, and selection can be read rather than reconstructed from execution. We build such organisms in an interactive 3D simulation of breeding creatures whose environment cycles through five conditions of food and hazard. Each creature carries eleven categories of written genes and acts by retrieving the sentences that fit its situation. Offspring inherit by recombining parental sentences within each category, with the language model blending the inherited text, and a closed-loop component writes new genes when no existing one covers a situation. Across 5,307 births, offspring gene text resembles parental text far more closely than randomly paired text (Cohen's d = 5.55, p ≈ 0). Changing the environment shifts behavior in all five measured dimensions (F = 223–1,009, p ≈ 0), while under stable conditions heritable variation persists for roughly 340 generations with nothing acting to preserve it. With mutation disabled so that gene wording is preserved, creatures whose genes tell them to flee leave 15% more offspring than creatures told to rally (d = 0.45), and 21% more when each carries two copies of every gene instead of one (d = 0.54), so selection acts on the readable text itself. A natural-language genome therefore works as a genetic system, giving artificial life researchers a substrate in which the genotype, its expression, and the targets of selection are all legible as text.