Behavioral Inheritance and Evolution in LLM-Controlled Agent Populations

Behavioral genetics need not be exclusive to biological organisms. Any computational agent whose behavior is specified in natural language can carry a readable genome that is inherited, selected, and evolved. Populations of such agents exhibit faithful behavioral inheritance, selection, and runtime evolution. Natural-language instruction corpora serve as the genetic substrate. Under the BEAR framework, each agent carries a per-entity instruction corpus that serves as its genome. Behavioral specifications are recombined across parents, mutated by the LLM, and expressed through context-aware retrieval. A closed-loop engine detects behavioral gaps and synthesizes new instructions under safety constraints, giving runtime evolution without retraining. An interactive 3D simulation validates these mechanisms, with eleven gene categories and epoch-driven environmental cycling across five conditions. Inheritance is faithful (per-gene cosine d = 5.55, p ≈ 0, 5,307 births). Epoch-driven behavioral shifts are significant across all five dimensions (F = 223–1,009, p ≈ 0). Heritable diversity is stable over 340.6 ± 24.9 generations. Action-tag instrumentation links inherited gene text directly to fitness. Under mutation rate zero, flee-strategy alleles produce 15% more offspring than rally-strategy carriers in single-allele mode (d = 0.45, p ≈ 3 × 10⁻¹⁸). The effect amplifies to 21% in diploid mode (d = 0.54, p ≈ 9 × 10⁻¹⁹).