Great post! Some clarification questions:
Just to confirm, inputs are sampled without replacement. So the optimizer agent can’t test different strategies on the same input?
The optimizer only gets the mean score from the judge, not the score per poem?
it gets per poem scores—i don’t let it select to regen on the exact same poems (I also did some runs where the optimizer could pick to rerun, but it didn’t seem to affect perf too much0
Great post! Some clarification questions:
Just to confirm, inputs are sampled without replacement. So the optimizer agent can’t test different strategies on the same input?
The optimizer only gets the mean score from the judge, not the score per poem?
it gets per poem scores—i don’t let it select to regen on the exact same poems (I also did some runs where the optimizer could pick to rerun, but it didn’t seem to affect perf too much0