Characterize the effect of relaxed acceptance on output diversity

Characterize how epsilon-threshold acceptance in HyCo changes the diversity of generated outputs relative to exact verification and the verifier’s original sampling distribution.

Background

HyCo’s default epsilon-threshold acceptance rule can accept drafted tokens that the verifier did not sample, so it does not preserve the verifier’s sampling distribution. The paper evaluates quality using judge scores and summary metrics, but these evaluations do not address whether the relaxed rule changes output diversity. Establishing this effect would clarify an important quality and distributional consequence of lossy acceptance.