Code World Models for Parameter Control in Evolutionary Algorithms

Researchers Camilo Chacón Sartori and Guillem Rodríguez Corominas have published a breakthrough paper demonstrating how large language models can learn to control evolutionary optimization algorithms. Their Code World Models (CWMs) approach extends LLM-synthesized Python programs from deterministic games to stochastic combinatorial optimization, enabling an LLM to create a simulator of an optimizer's dynamics by observing suboptimal trajectories. The synthesized simulator then enables greedy planning to select optimal mutation parameters at each step, achieving performance within 6% of theoretically optimal policies without ever seeing optimal trajectories.

The technical results are striking: on deceptive Jump-k problems where all adaptive baselines failed with 0% success rates, CWM-greedy achieved 100% success without oracle knowledge of gap parameters. The approach demonstrated superior sample efficiency (200 offline trajectories vs. 500 online episodes for DQN), better success rates (100% vs. 58%), and stronger generalization (78% vs. 0% for k=3 problems). On NK-Landscape problems with no closed-form models, CWM-greedy outperformed all baselines across fifteen instances (36.94 vs. 36.32 average fitness, p<0.001) when prompts included empirical transition statistics, showing robust synthesis across five independent runs.