A new AI framework can design entire optimization algorithms from scratch, not just tweak one piece of a human-built template.
Researchers built ATLAS, a system that lets large language models write full algorithms for combinatorial optimization problems from the ground up. It only fixes the format for inputs and outputs; the LLM decides everything else, including which techniques to use, how components interact, and how the whole thing is structured. That open-ended freedom means more chances for the AI to produce broken code or converge on one narrow design, so ATLAS automatically flags execution, interface, and feasibility failures, repairs them, and keeps a diverse archive of algorithms drawn from different regions of the design space. A three-layer search then polishes the strongest designs, gives other promising ones room to improve, and recombines pieces across different algorithms.
Across four NP-hard problems, ATLAS beat other AI-driven algorithm-design methods and came close to matching strong human-designed algorithms - a notable result, since most AI code-generation tools are still boxed into filling one slot of a pre-built scaffold. It also produced multiple distinct, comparably strong algorithms in a single run, built around different construction and search strategies, rather than converging on one design.
"Came close to matching" is the operative phrase: NP-hard problems are exactly the terrain where decades of human tuning are hardest to beat, and ATLAS still didn't.