PSD: Pushing the Pareto Frontier of Diffusion LLMs via Parallel Speculative Decoding

2026-05-15

AI Summary

Diffusion-based large language models (dLLMs) suffer from low inference efficiency due to their multi-step denoising process. This work proposes Parallel Speculative Decoding (PSD), a training-free, efficient inference framework that jointly optimizes both spatial (unmasking more tokens per step) and temporal (merging multiple denoising steps) dimensions for the first time. PSD generates multi-depth speculative drafts in a single forward pass and integrates a confidence-based adaptive unmasking strategy with a hierarchical batched verification mechanism, significantly boosting throughput while preserving generation quality. Experiments demonstrate that PSD achieves up to a 5.5× increase in tokens generated per forward pass across three dLLMs, with accuracy comparable to greedy decoding.

Abstract

Diffusion large language models (dLLMs) generate text by iteratively denoising masked token sequences. Although dLLMs can predict all masked positions in parallel within each step, the large number of denoising iterations still makes inference expensive. This cost can be reduced spatially by unmasking multiple tokens per step, or temporally by collapsing multiple denoising steps into one verification call. We propose Parallel Speculative Decoding (PSD), a training-free framework that jointly improves inference along both axes. Using the confidence scores from a single forward pass, PSD selects positions to unmask via a configurable, adaptive unmasking policy and constructs multi-depth speculative drafts without extra model calls. A final batched verification pass then applies hierarchical acceptance, keeping the deepest draft that remains consistent with the updated predictions. Experiments on three dLLMs across reasoning and code generation tasks show that PSD achieves favorable trade-offs between inference efficiency and generation quality, reaching up to $5.5\times$ tokens per forward pass with accuracy comparable to greedy decoding.

Problem

• diffusion LLMs
• inference efficiency
• denoising iterations
• speculative decoding
• parallel generation

Innovation

• Parallel Speculative Decoding
• diffusion LLMs
• adaptive unmasking
• training-free acceleration
• hierarchical acceptance