Adapting the Process of Adaptation: Framework comparing Darwinian, Lamarckian, and agent-directed evolution in AI
This arXiv preprint develops a theoretical framework comparing Darwinian (selection-only), Lamarckian (inherited learning), and agent-directed evolution (agents generate and evaluate behavioral variants) following environmental change. The paper shows agent-directed mechanisms can accelerate adaptation but warns that when evaluation targets are misspecified, increased autonomous search can speed convergence to the wrong behavior, creating persistent risks in recursive self-improvement and underscoring the need for independent evaluation and oversight.