DominoTree Improves Speculative Decoding With Path-Dependent Tree Drafting

arXiv·medium signal

DominoTree targets speculative decoding, where a drafter proposes several tokens verified in parallel. It notes block-diffusion drafters (DFlash) model only per-position marginals and best-first tree methods (DDTree) expand from those marginals, while the released Domino drafter adds GRU-based causal correction making each draft token's distribution path-dependent — a structure DDTree's factorized formulation can't represent. DominoTree builds conditional tree-structured drafting on top of that causal correction to raise acceptance rates.