Adversarial Preference

A training objective that makes the draft model deliberately deviate from the target model at critical decision points. By triggering verification rejection and resampling, it causes key tokens to be regenerated by the target model under low-noise conditions.