We open-sourced EvoAgent-NO at Chips & Superintelligence Labs.
It pretrains neural operators with zero data. A small Python program poses physics problems, an exact GPU solver answers them, and a Fourier neural operator trains on the answers. When the operator stops improving, LLM agents rewrite the program that poses the problems, and a rewrite is kept only if a short training trial shows the operator learns more from it. Self-play inside each round is optional.
It is for when you can simulate something but do not have a dataset: a fast surrogate you pretrain first and fine-tune on the few real runs you have, a rough guess at the problem distribution that the agents shape, or a solver set up for one regime that the agents extend to the rest. At the lab we care about the chip side of that: thermal maps, IR drop, dopant diffusion and lithography imaging.
The write-up on the lab's site has the architecture, problems the solver actually posed, a real rewrite by Claude and the use cases in more detail: EvoAgent-NO: Agents that evolve what a physics computer learns.