That's not how it works. Look at AlphaEvolve. The model generates hypotheses and designs experiments, and the results of those experiments are fed into the next round, with notable results percolated up to humans for refinement.