School of Computer Science Seminar Professor Lars Kotthoff

Professor Lars Kotthoff will present LLaMEA-CFG: Code-Feature Guidance for Automated Algorithm Design

Abstract: Large language models (LLMs) have advanced automated algorithm design (AAD) by generating complete optimization algorithms from natural-language prompts. Evolutionary frameworks such as LLaMEA explore this design space effectively, but rely solely on fitness feedback, leaving information about the structure of generated code unused. We propose Code-Feature Guidance (CFG), which extracts graph-theoretic and complexity features from the code of generated algorithms, trains a surrogate model over an archive of evaluated solutions, and applies SHAP feature-attribution to identify the most influential feature. This is translated into a natural-language mutation instruction that steers subsequent LLM-based code generation without restricting expressivity. Integrating CFG into LLaMEA (LLaMEA-CFG), we show it reaches comparable performance faster than vanilla LLaMEA in a controlled experiment, and achieves competitive or superior performance against state-of-the-art AAD methods on the MA-BBOB suite from the GECCO-MA-BBOB competition, demonstrating that signals derived from code can effectively bias LLM-driven algorithm evolution.

Bio: Lars Kotthoff is a professor of Computer Science at the University of St Andrews, where he holds the Johann and Gaynor Rupert chair in Artificial Intelligence. He was a visiting professor at Sorbonne Université and held an endowed chair at the University of Wyoming, where he maintains an adjunct position. Lars has held postdoctoral appointments at the University of British Columbia and University College Cork, after obtaining a PhD from the University of St Andrews. He has made contributions to automated machine learning, algorithm selection, applications of AI in materials science, and various other areas that he found interesting. Lars is one of the core contributors to the mlr3 machine learning framework.