DARTree Extends AR Correction Heads From Draft Chains to Trees, Hitting 9.73x Lossless Speedup

arXiv 2608.13524·medium signalXBlueskyLinkedInCopy link

Diffusion drafters propose whole token blocks in parallel but their position-wise distributions are marginal, not conditioned on tokens chosen along each draft path. DARTree is a training-free method that builds a fixed-width candidate tree by expanding and scoring all nodes at each depth in one batch, then applies best-first pruning to select the verification tree — decoupling AR-head inference from sequential heap operations. Across seven math, code and chat benchmarks it leads average acceptance length and speedup in all four model–temperature configurations, accepting up to 12.97 tokens per verification round (98.6% more than DFlash, 27.9% more than Domino) for up to 9.73x lossless speedup over locally measured autoregressive decoding.