[KongchangAI]
ConceptDraft Tree

Draft Tree

An extension of the speculative decoding structure that expands a single candidate token chain into a tree of multiple parallel candidate paths, allowing the target model to verify multiple candidate nodes in a single forward pass and theoretically increasing the average accepted length without additional target model calls.

Timeline (last 90 days)

Sep 22

Existing draft tree construction methods rely on the marginal probability at each position to rank candidate nodes, ignoring which parent node a candidate extends from, which is especially detrimental for semi-autoregressive drafters.

Unverified50%

All Facts (1)

Source Articles