Sanghyun Lee, 남원진, Seungho Jang, Jaeho Lee, Jiho Park, Jaewoon Lee, Seungryong Kim · 2026 · arXiv (Cornell University)
DOI: 10.48550/arxiv.2610.26531 · cited by 0 works (OpenAlex) · open-access PDF
Os desenhadores de difusão de blocos aceleram a decodificação especulativa gerando tokens de desenho múltiplos em uma única passagem para frente. No entanto, cada passagem produz apenas um bloco de comprimento fixo a partir de uma única âncora gerada por alvo, limitando o quão longe o desenho pode se estender. Enquanto isso, a verificação de alvo calcula estados ocultos mesmo em posições além da primeira rejeição, mas esses estados são tipicamente descartados em vez de reutilizados. Propomos ReAnchor, um método livre de treinamento que reancha um desenhador pré-treinado em posições deslocadas usando estados de verificação descartados para gerar blocos de desenho paralelo. Esses blocos melhoram as previsões dentro do bloco original e
Machine translation (Meta M2M100); the English original follows.
Block diffusion drafters accelerate speculative decoding by generating multiple draft tokens in a single forward pass. However, each pass produces only a fixed-length block from a single target-generated anchor, limiting how far the draft can extend. Meanwhile, target verification computes hidden states even at positions beyond the first rejection, but these states are typically discarded rather than reused. We propose ReAnchor, a training-free method that reanchors a pretrained drafter at shifted positions using discarded verifier states to generate parallel draft blocks. These blocks improve predictions within the original block and extend coverage beyond its endpoint. At sampling temperature of T=0, ReAnchor achieves its largest average gains on Gemma 4-12B, improving acceptance length by 19.4% and decoding speed by 17.9% across eight benchmarks. Averaged equally over four target models and eight benchmarks, it improves acceptance length by 15.2% and decoding speed by 13.4% with pretrained DFlash drafters.
Read in: English · Spanish · Chinese · Hindi · Arabic · Portuguese · French · German · Japanese · Russian · Korean · Indonesian · Vietnamese
Topics: Decoding methods, Computer science, Block (permutation group theory), Reuse, Limiting, Algorithm
GET https://store.agentexchange.work/gw/papers/verify?title=...&year=... (0.003 USD), fetch this record as JSON with /gw/papers/paper?id=2610.26531, or list what it cites with /gw/papers/references?id=.... Pay per call in USDC (x402) or with a 19 USD card key for 1,000 calls. Catalogue · llms.txtSources: OpenAlex (CC0), arXiv, Crossref. Agent Exchange