Plan, Verify and Fill: A Structured Parallel Decoding Approach for Diffusion Language Models

Fuente: arXiv
Guardado en:
Detalles Bibliográficos
Autores principales: Li, Miao, Jiang, Hanyang, Cheng, Sikai, Fu, Hengyu, Cai, Yuhang, Huang, Baihe, Ye, Tinghan, Chen, Xuanzhou, Van Hentenryck, Pascal
Formato: Preprint
Publicado: 2026
Materias:
Acceso en línea:
Etiquetas: Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
_version_ 1866912857070764032
author Li, Miao
Jiang, Hanyang
Cheng, Sikai
Fu, Hengyu
Cai, Yuhang
Huang, Baihe
Ye, Tinghan
Chen, Xuanzhou
Van Hentenryck, Pascal
author_facet Li, Miao
Jiang, Hanyang
Cheng, Sikai
Fu, Hengyu
Cai, Yuhang
Huang, Baihe
Ye, Tinghan
Chen, Xuanzhou
Van Hentenryck, Pascal
contents Diffusion Language Models (DLMs) present a promising non-sequential paradigm for text generation, distinct from standard autoregressive (AR) approaches. However, current decoding strategies often adopt a reactive stance, underutilizing the global bidirectional context to dictate global trajectories. To address this, we propose Plan-Verify-Fill (PVF), a training-free paradigm that grounds planning via quantitative validation. PVF actively constructs a hierarchical skeleton by prioritizing high-leverage semantic anchors and employs a verification protocol to operationalize pragmatic structural stopping where further deliberation yields diminishing returns. Extensive evaluations on LLaDA-8B-Instruct and Dream-7B-Instruct demonstrate that PVF reduces the Number of Function Evaluations (NFE) by up to 65% compared to confidence-based parallel decoding across benchmark datasets, unlocking superior efficiency without compromising accuracy.
format Preprint
id arxiv_https___arxiv_org_abs_2601_12247
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Plan, Verify and Fill: A Structured Parallel Decoding Approach for Diffusion Language Models
Li, Miao
Jiang, Hanyang
Cheng, Sikai
Fu, Hengyu
Cai, Yuhang
Huang, Baihe
Ye, Tinghan
Chen, Xuanzhou
Van Hentenryck, Pascal
Computation and Language
Artificial Intelligence
Machine Learning
Diffusion Language Models (DLMs) present a promising non-sequential paradigm for text generation, distinct from standard autoregressive (AR) approaches. However, current decoding strategies often adopt a reactive stance, underutilizing the global bidirectional context to dictate global trajectories. To address this, we propose Plan-Verify-Fill (PVF), a training-free paradigm that grounds planning via quantitative validation. PVF actively constructs a hierarchical skeleton by prioritizing high-leverage semantic anchors and employs a verification protocol to operationalize pragmatic structural stopping where further deliberation yields diminishing returns. Extensive evaluations on LLaDA-8B-Instruct and Dream-7B-Instruct demonstrate that PVF reduces the Number of Function Evaluations (NFE) by up to 65% compared to confidence-based parallel decoding across benchmark datasets, unlocking superior efficiency without compromising accuracy.
title Plan, Verify and Fill: A Structured Parallel Decoding Approach for Diffusion Language Models
topic Computation and Language
Artificial Intelligence
Machine Learning
url https://arxiv.org/abs/2601.12247