To answer your second question, rebuttals for both sides are rolled out at the same time per turn, meaning that Alice and Bob both see the debate as it exists at turn as they generate their responses for turn . Since our “constructive” speeches consist of the proposer’s solution and the critic’s stance (which is generated with the proposer’s solution in context), and since both of these positions are both presented to all parties prior to the beginning of the round, we treat all actual rounds of debate as rebuttals. So our results effectively say that only the proposer’s solution and critic’s stance are necessary to produce the improvement we see in debate.
To answer your first question, we wrote prompts for each setting, although we limited our iterations so as not to overfit. The things that we implemented are relatively basic coaching: solve step-by-step, come up with a list of adversarial examples and then walk through them one-by-one, try to solve independently first without using the proposer’s answer, etc. I will add the specific prompts to the preprint in the next few days.
To answer your second question, rebuttals for both sides are rolled out at the same time per turn, meaning that Alice and Bob both see the debate as it exists at turn
To answer your first question, we wrote prompts for each setting, although we limited our iterations so as not to overfit. The things that we implemented are relatively basic coaching: solve step-by-step, come up with a list of adversarial examples and then walk through them one-by-one, try to solve independently first without using the proposer’s answer, etc. I will add the specific prompts to the preprint in the next few days.