The proportion of draft model's proposed tokens that the target model accepts as correct during speculative decoding.