Debate May Help AI Models Converge on Truth