ATLAS by researchers autonomously designs algorithms without scaffolding

A team of researchers from RMIT University, the University of Birmingham, and the University of Melbourne has unveiled ATLAS, a groundbreaking framework that enables large language models (LLMs) to autonomously design complete algorithms without human scaffolding. Published on arXiv (arXiv:2608.15546), ATLAS introduces an embedding-guided quality-diversity search mechanism that allows LLMs to explore and refine algorithmic designs across a vast, scaffold-free search space. Unlike prior approaches that optimize limited components within predefined structures, ATLAS starts from minimal input/output specifications and lets the LLM restructure components, interactions, and control flow entirely on its own.

The framework employs a three-layer search process: error detection and repair for invalid candidates, similarity-based archive management to maintain diversity across embedding-space regions, and cross-region synthesis to recombine components. In evaluations across four NP-hard combinatorial optimization problems, ATLAS not only outperformed several state-of-the-art component-synthesis methods but also remained competitive with strong human-designed algorithms. Notably, ATLAS can return multiple high-performing algorithms from distinct design regions in a single run, often revealing novel design patterns such as alternative global-search backbones or primary construction strategies.