rejection sampling
OUTPUT
NeurIPS 2026 is in
p(x
i
), the target model's probability
q(x
i
), the draft model's probability
a(x
i
) = min(1, p/q), the accept probability