Researchers have developed a new algorithm called ATLAS that allows large language models (LLMs) to design complex algorithms from scratch without relying on pre-defined templates or scaffolds. This approach, known as embedding-guided quality-diversity search, enables the LLMs to explore a much larger space of possible solutions and can lead to more innovative designs. In experiments, ATLAS outperformed several state-of-the-art methods and was competitive with human-designed algorithms on four NP-hard problems. The researchers also found that ATLAS can generate multiple high-quality algorithms from different regions of the solution space. --- Why it matters: This matters because it shows that large language models can be used to explore complex design spaces in a more efficient and effective way, potentially leading to breakthroughs in fields like optimization, machine learning, and artificial intelligence. The ability to design novel algorithms without relying on pre-existing templates or scaffolds could also lead to new applications and innovations. Source: https://arxiv.org/abs/2608.15546