AI-generated code beats humans at link prediction in networks

In a new preprint on arXiv, researchers Vlaskin and Altmann explore automated code evolution for the classic problem of link prediction in complex networks. They applied a combination of large language models and genetic algorithms to evolve algorithms that predict missing or future connections in networks (e.g., social, biological, or infrastructure networks). Despite training on limited data, the evolved algorithms significantly outperformed state-of-the-art human-designed methods across 580 diverse networks, achieving an average AUC of 0.915 versus 0.783.

The discovered methods not only matched known human strategies but introduced key innovations in how node and link features are selected and combined. Notably, the evolved algorithms are computationally efficient, scaling to networks with millions of links. The authors argue that this work demonstrates a broader potential for AI-driven algorithmic innovation and scientific discovery, where code-evolution systems can autonomously generate novel, high-performance solutions in domains traditionally dominated by human expertise.