Block draftors such as DFlash generate a block of future tokens in a single forward pass, while draft trees like DDTree verify the prefixes that receive the highest rank according to the product of the draftor’s per‑position distributions. Under tree verification, the expected number of accepted tokens equals the probability mass covered by the tree. We show that the product misallocates this mass: when a continuation can stem from several futures, the product splices them into chimeras—prefixes stitched together from different futures that the target model never produces—yet allocates them the same budget as the genuine futures. Except for exact ties, any product‑ranked tree cannot hold two futures that disagree in $m$ positions without generating at least $2^{m-1}-1$ chimeras. In practice, chimeras constitute 27%–33% of a 64‑node tree.
To address this, we propose Multi‑Future Drafting (MFD). MFD restores the coupling with a latent variable instead of the chain‑rule product: $K$ learned mode embeddings are fed to the draftor’s input, allowing it to draft $K$ coherent blocks in the same pass. A mixture DDTree then lets the verifier pick the future actually taken by the target, losslessly, using only $K+1$ extra vectors. Formally, MFD models $$ P(y\mid x) = \sum_{k=1}^{K} P(z=k\mid x)\,P(y\mid x, z=k) $$ where $z$ is a latent mode variable.
Evaluated on four Qwen targets ranging from 4B to 27B parameters with batch size 1, MFD achieves up to 7.8× speed‑up, 5%–8% faster than the best single‑mode tree trained under identical conditions. On Qwen3‑4B and Qwen3‑8B, MFD is 9%–16% faster than the released DARTree trained on the same prompts.
Review: By explicitly modeling multiple futures and introducing a latent variable, MFD eliminates the chimera issue inherent to product‑based ranking, delivering substantial efficiency and robustness gains for tree‑verified speculative decoding. This work highlights a promising direction for accelerating inference in large language models.