> (1) What's the value add of the LLM here? Does it substantially reduce the number of evaluations necessary to converge? If so, how? I thought that stochastic

> (1) What's the value add of the LLM here? Does it substantially reduce the number of evaluations necessary to converge? If so, how? I thought that stochastic (random) gradient descent and LLM were converging much quicker than genetic programming. Definitely much quicker than random search.