AI That Creates Truly Novel Algorithms: A Review of Learning-Based Algorithm Discovery, Automated Evaluation, and Evolutionary Search

This paper reviews recent advances in artificial intelligence systems capable of discovering and optimizing algorithms. It examines AlphaTensor, AlphaDev, FunSearch, and AlphaEvolve, focusing on their approaches to algorithm discovery, search, evaluation, and optimization. The paper also proposes the Verified Algorithm Discovery Pipeline (VADP), a conceptual framework combining candidate generation, automated evaluation, verification, and iterative refinement. VADP is a proposed framework and has not been experimentally implemented or validated in this work. The paper discusses current limitations, verification challenges, and future research directions in AI-driven algorithm discovery.