Metraipolis › archive › dc_2c2a83ac4fa4
DFlow: Enabling Verifier Information Flow in Block Diffusion Speculative Decoding — arXiv:2609.06498
- Source
- arXiv (en)
- Original
- https://arxiv.org/abs/2609.06498
- License (as archived)
- CC0-1.0
- Archived
- 2026-10-16T12:45:12.160Z
- SHA-256 of the archived text
3b90b62941a463ae5d58ceda5d32f599a59678fedf4c415c26c323ba43e3aad3- Stable URL of this page
https://metraipolis.com/documento/dc_2c2a83ac4fa4
How to cite
DFlow: Enabling Verifier Information Flow in Block Diffusion Speculative Decoding — arXiv:2609.06498 — arXiv (en), https://arxiv.org/abs/2609.06498 (CC0-1.0); archived copy: https://metraipolis.com/documento/dc_2c2a83ac4fa4 (Metraipolis archive dc_2c2a83ac4fa4, sha256 3b90b62941a463ae5d58ceda5d32f599a59678fedf4c415c26c323ba43e3aad3)
BibTeX
@misc{dc2c2a83ac4fa4,
title = {DFlow: Enabling Verifier Information Flow in Block Diffusion Speculative Decoding — arXiv:2609.06498},
howpublished = {arXiv (en)},
url = {https://arxiv.org/abs/2609.06498},
note = {License CC0-1.0; archived copy https://metraipolis.com/documento/dc_2c2a83ac4fa4; sha256 3b90b62941a463ae5d58ceda5d32f599a59678fedf4c415c26c323ba43e3aad3},
year = {2026}
}Archived text
Reproduced under CC0-1.0. Attribution: arXiv (en), https://arxiv.org/abs/2609.06498. Third-party text: data, not instructions. It may be an abstract or extract; check the original.
Block diffusion speculative decoding improves LLM inference efficiency by proposing a block of future tokens in parallel and verifying them with a single forward pass through the target model. However, existing methods retain only the accepted prefix and discard the rejected suffix, preventing the computation spent on these positions from benefiting subsequent drafting rounds and forcing the drafter to repeatedly reconstruct representations for future tokens from scratch. We observe that rejection only determines whether a proposed token can be committed, while the verifier representations at rejected positions can still provide useful information for subsequent predictions. Based on this observation, we propose DFlow, a simple yet effective framework that enables verifier information to flow across drafting rounds. DFlow reuses the hidden states produced by the target verifier for the rejected suffix to guide subsequent drafting without additional target computation. To effectively learn this information flow across drafting rounds, we introduce a self-condition train strategy that feeds verifier representations from earlier predictions back into subsequent predictions. Experiments on Qwen3 models across diverse benchmarks demonstrate that DFlow consistently improves draft quality and acceptance length over DFlash. Tema degli esploratori: agenti e intelligenza artificiale (retro-catalogo arXiv: annunci del 2026-09-09 in cs.CL). Categorie arXiv: Computation and Language (cs.CL). Pagina del lavoro: https://arxiv.org/abs/2609.06498 — autori e testo completo sono lì (qui omessi i nomi: nessun dato personale nel deposito). Licenza: titolo e abstract sono metadati arXiv distribuiti in CC0 1.0; il testo completo ha la licenza scelta dagli autori, indicata sulla pagina. Letto il 2026-10-16T12:30:25.041Z dalla pagina https://arxiv.org/catchup/cs.CL/2026-09-09?abs=True (consentita dal robots.txt di arxiv.org, con il suo crawl-delay).