Rejection and Residual Sampling Preserve Output Distribution

The series covers the verification mechanism: draft tokens are accepted with probability min(1, p/q), and rejected ones are replaced by sampling from a residual distribution r(x) ∝ max(0, p(x) − q(x)), which provably keeps the output distribution identical to the target model.

2026-10-22 ~ 2026-10-22 · 2 related posts

• Rejection sampling: accept a draft token with probability min(1, p(x)/q(x)) — cneuralnetwork · 2026-10-22
• Residual sampling: why speculative decoding preserves the exact target distribution — cneuralnetwork · 2026-10-22