Speculative decoding needs better benchmarks than acceptance rate

has a formal guarantee: with , the final output stays the same as the larger target model. The quality of the smaller draft model does not decide correctness.

A weak draft model gets rejected more often, so it can reduce the speed benefit, but it does not change what the target model would say. This makes it different from , where speed or size gains can come with a quality tradeoff.

The is directly tied to the gap between the draft model and target model s: it equals 1 minus the . The practical question is why a lossless speed method is not the default in more local setups, and what real speed gains different draft-target pairs actually deliver.

Key points

  • uses a smaller draft model to propose tokens before a larger target them.
  • keeps the final output aligned with the target model.
  • A poor draft model mainly hurts speed, not output quality.
  • equals 1 minus between the two models' s.
  • Benchmarks should report real speed and cost impact, not only .
Read original