Behavioral Inheritance and Evolution in LLM-Controlled Agent Populations

We demonstrate that populations of LLM-controlled agents can exhibit
behavioral inheritance, selection, and evolution using natural-language
instruction corpora as a genetic substrate. Agents governed by the BEAR
framework carry per-entity instruction corpora that serve as genomes.
Behavioral specifications are recombined across parent agents, mutated
by the LLM, and expressed through context-aware retrieval. A closed-loop
engine detects behavioral gaps and synthesizes new instructions under
safety constraints, implementing runtime evolution without retraining.

We validate these mechanisms through an interactive 3D simulation with
eleven gene categories and epoch-driven environmental cycling across
five conditions. Agents demonstrate faithful behavioral inheritance
(per-gene cosine d = 5.55, p ≈ 0, 5,307 births), significant
epoch-driven behavioral shifts across all five dimensions
(F = 223–1,009, p ≈ 0), and stable heritable diversity over
340.6 ± 24.9 generations.

Action-tag instrumentation directly links inherited gene text to
fitness outcomes: under mutation rate zero, flee-strategy alleles
produce 15% more offspring than rally-strategy carriers in
single-allele mode (d = 0.45, p ≈ 10⁻¹⁸), with the effect amplified
to 21% in dual-allele mode (d = 0.54, p ≈ 10⁻¹⁹).

Together, these results establish natural-language behavioral genetics
as a viable substrate for inheritance, selection, and runtime evolution
in LLM-controlled agent populations. The action-marker decay observed
at the default mutation rate identifies a characterized failure mode
of LLM-mediated text operations on literal tokens, motivating future
mutation operators that preserve actionable substrate.