Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

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

Miao Li, Hanyang Jiang, Sikai Chen, Hengyu Fu, Yuhang Cai, Baihe Huang, Tinghan Ye, Xuanzhou Chen, Pascal Van Hentenryck

TL;DR

Diffusion Language Models enable parallel, global decoding but current strategies are largely reactive. PVF introduces a training-free Plan-Verify-Fill approach that explicitly builds a high-level structural skeleton via planning tokens and verifies them with an impact-set before filling details, using an autoregressive fallback when planning stalls. The method yields major efficiency gains, reducing the Number of Function Evaluations ($NFE$) by up to 65% on several benchmarks while maintaining accuracy, and demonstrates model-agnostic robustness across LLaDA-8B-Instruct and Dream-7B-Instruct. PVF achieves these results through a dual-route architecture, automated discovery of planning tokens via structural distillation, and a safety-first verification regime, with potential future work in training-aware structure alignment to further boost speed and reliability.

Abstract

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.

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

TL;DR

Diffusion Language Models enable parallel, global decoding but current strategies are largely reactive. PVF introduces a training-free Plan-Verify-Fill approach that explicitly builds a high-level structural skeleton via planning tokens and verifies them with an impact-set before filling details, using an autoregressive fallback when planning stalls. The method yields major efficiency gains, reducing the Number of Function Evaluations () by up to 65% on several benchmarks while maintaining accuracy, and demonstrates model-agnostic robustness across LLaDA-8B-Instruct and Dream-7B-Instruct. PVF achieves these results through a dual-route architecture, automated discovery of planning tokens via structural distillation, and a safety-first verification regime, with potential future work in training-aware structure alignment to further boost speed and reliability.

Abstract

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.
Paper Structure (49 sections, 18 equations, 8 figures, 3 tables, 1 algorithm)

This paper contains 49 sections, 18 equations, 8 figures, 3 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Parallel Decoding
  4. Planning Tokens
  5. Hierarchy-aware decoding for DLMs
  6. Preliminary
  7. Reverse Process (Denoising):
  8. Confidence-based Parallel Decoding:
  9. Semi-Autoregressive Blockwise Decoding
  10. Inducing Hierarchical Structure with Planning Tokens
  11. The Safety of Content-Neutral Anchors.
  12. Restricting Search Space via Structural Constraints.
  13. Automated Discovery via Structural Distillation
  14. Methodology: Plan-Verify-Fill (PVF) Decoding
  15. Overview of the PVF Pipeline
  16. ...and 34 more sections

Figures (8)

  • Figure 1: Overview of the Plan--Verify--Fill (PVF) decoding pipeline.
  • Figure 2: Ablation on GSM8K and HumanEval comparing lower-confidence commits of planning tokens versus random tokens (i.e., without prioritizing planning tokens). Across confidence bins, prioritizing planning tokens consistently yields faster decoding and improved accuracy; points closer to the upper-right indicate better performance on both axes.
  • Figure 3: Overview of the Planning Route
  • Figure 4: Overview of the AR Fallback Route
  • Figure 5: Ablation study on the full GSM8k dataset evaluating the contribution of each PVF component. Accuracy scores are displayed above each bar to confirm they remain comparable across methods.
  • ...and 3 more figures