Can AI agents discover better machine learning algorithms by learning from their mistakes?

MLEvolve uses multi-agent collaboration and a memory system to let LLM agents iteratively discover new machine learning algorithms. Unlike prior systems that forget past attempts and can't share insights across search branches, MLEvolve connects discoveries via a graph structure and stores reusable patterns. On MLE-Bench, it achieves top rankings while cutting runtime in half, and outperforms specialized algorithm discovery tools like AlphaEvolve on mathematical tasks. Code is released.