Towards a Universal Algorithm Constructor: Evolutionary Computation in the Age of LLMs
Date:
Abstract
What if evolutionary computation is not just a tool for optimization, but a design engine for algorithms themselves?
In this keynote, I argue that we are witnessing a shift from hand-crafted metaheuristics to automated algorithm construction, with evolutionary principles providing the backbone and Large Language Models and other generative AI techniques acting as high-level variation operators. Rather than replacing EC, LLMs amplify its expressive power, enabling search directly in program space.
Using the LLaMEA framework and its extensions (in-the-loop HPO, MAP-elites, Novelty search and others), I will show how evolutionary loops can generate competitive algorithms across continuous, combinatorial, and real-world engineering problems. But raw performance is not enough. I will discuss how mutation control, prompt design, and Code Evolution Graphs expose the internal dynamics of LLM-driven search and why rigorous benchmarking is essential to separate genuine progress from hype.
The key question is no longer whether we can automate algorithm design, but how to better control, understand and generalize it to new domains. Are we building specialized optimizers, or the first steps toward a universal algorithm constructor?