← News

ResearchSources: arXiv · 19 Nov 2026

DBloom post-training extends block-diffusion drafters past their ceiling, arXiv paper reports

A new arXiv paper argues that high acceptance rates can hide unrealized speed-up in speculative decoding, and proposes a short post-training curriculum to expand the drafter's block size.

In plain words

Speculative decoding uses a small draft model to guess several tokens at once, which a larger model then checks in a single pass. The paper's authors found cases where the larger model accepts every guessed token, so the drafter runs out of trained capacity before verification fails. Their fix, DBloom, continues training the drafter to handle a longer block of tokens.

What was announced

The authors, writing on arXiv, call the phenomenon stranded speed-up: in many cycles the target model accepts the entire draft block, so the drafter exhausts its trained horizon before verification ever fails. They argue that judging a drafter by mean committed length alone hides this bottleneck, and recommend inspecting the acceptance histogram — where the problem appears as a spike in the ceiling bin — before spending more training compute. They also caution that simply widening the block at inference can hurt acceptance for drafters with bidirectional attention, because once the block exceeds its training size the drafter's distribution shifts even at early positions.

As a remedy, the paper introduces DBloom, a short curriculum that post-trains the drafter at its original block size before expanding it, with emphasis on newly exposed positions. The authors continued training pretrained DFlash and DFlare drafters from a block size of 16 to 24 on Qwen3-8B and Qwen3-4B. They report that, compared with the original B16 drafters, this raises per-prompt committed length by up to 1.7 tokens and decode speed-up by a median of 10.2%, versus 5.7% without the continuation, rising to 11.9% on high-ceiling benchmarks. On Gemma-4-12B-IT and GPT-OSS-20B, expansion raised committed length on every benchmark, and against the JetSpec tree drafter, DBloom showed higher committed length by point estimate on every benchmark at tree budgets up to 64 nodes.

What it does not show

  • The evidence is the paper's abstract only; full methodology, experimental setup and peer review status are not covered by the supplied excerpt.
  • Reported gains are the authors' own measurements and have not been independently confirmed.
  • The comparison with JetSpec is stated only as a point estimate, without significance testing described in the excerpt.

Sources

Source-support review completed 20 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 20 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.