Large language models generate text one token at a time, a slow process. accelerates this by using a small 'drafter' model to propose multiple tokens at once, which a larger 'target' model then verifies in parallel. But training the drafter to make better proposals traditionally requires sequential collection of full responses, which is inefficient.
SpecOPD breaks this dependency. It collects training data from the drafter's own guesses at independent branch points, allowing supervision to happen in parallel across many branches. Shared prefix computations and tree-structured attention further reduce cost.
SpecOPD achieved up to 41.45% more accepted tokens per proposal and up to 3.56 times faster generation than standard autoregressive decoding (one token at a time). This enables faster deployment of large models without sacrificing output quality.