![[2602.22260] Code World Models for Parameter Control in Evolutionary Algorithms](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.22260] Code World Models for Parameter Control in Evolutionary Algorithms
Summary
This paper explores the use of Code World Models (CWMs) to enhance parameter control in evolutionary algorithms, demonstrating significant performance improvements over traditional methods.
Why It Matters
The research addresses a critical challenge in evolutionary algorithms by leveraging large language models to optimize parameter control, potentially leading to more efficient and effective optimization strategies. This could have broad implications for fields relying on evolutionary algorithms, such as robotics and AI.
Key Takeaways
- CWMs can predict optimizer dynamics, improving parameter control.
- CWM-greedy outperforms traditional methods in complex scenarios.
- The approach shows robustness across multiple independent trials.
Computer Science > Machine Learning arXiv:2602.22260 (cs) [Submitted on 25 Feb 2026] Title:Code World Models for Parameter Control in Evolutionary Algorithms Authors:Camilo Chacón Sartori, Guillem Rodríguez Corominas View a PDF of the paper titled Code World Models for Parameter Control in Evolutionary Algorithms, by Camilo Chac\'on Sartori and Guillem Rodr\'iguez Corominas View PDF HTML (experimental) Abstract:Can an LLM learn how an optimizer behaves -- and use that knowledge to control it? We extend Code World Models (CWMs), LLM-synthesized Python programs that predict environment dynamics, from deterministic games to stochastic combinatorial optimization. Given suboptimal trajectories of $(1{+}1)$-$\text{RLS}_k$, the LLM synthesizes a simulator of the optimizer's dynamics; greedy planning over this simulator then selects the mutation strength $k$ at each step. On \lo{} and \onemax{}, CWM-greedy performs within 6\% of the theoretically optimal policy -- without ever seeing optimal-policy trajectories. On \jump{$_k$}, where a deceptive valley causes all adaptive baselines to fail (0\% success rate), CWM-greedy achieves 100\% success rate -- without any collection policy using oracle knowledge of the gap parameter. On the NK-Landscape, where no closed-form model exists, CWM-greedy outperforms all baselines across fifteen independently generated instances ($36.94$ vs.\ $36.32$; $p<0.001$) when the prompt includes empirical transition statistics. The CWM also outperforms DQ...
