dParallel: Learnable Parallel Decoding for dLLMs

2025-09-30

AI Summary

Diffusion large language models (dLLMs) hold promise for parallel decoding, yet existing methods require hundreds of iterative steps to ensure generation quality, severely undermining their efficiency advantage. This work proposes dParallel—a learnable parallel decoding framework—that identifies slow convergence of mask-token ordering confidence as the key bottleneck. To address this, we introduce certainty-forcing distillation, a novel training strategy integrating trajectory consistency constraints with a parallel confidence-boosting mechanism, significantly accelerating confidence convergence. On LLaDA-8B-Instruct, dParallel reduces decoding steps on GSM8K from 256 to 30 (8.5× speedup) and on MBPP from 256 to 24 (10.5× speedup), without performance degradation. To our knowledge, this is the first work achieving high-accuracy parallel sampling in dLLMs within merely tens of steps, establishing a new paradigm for efficient text generation.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: (Large) Language Models
• Search and Optimization: Sampling/Simulation-based Search

Application Category

• Search and Retrieval-Augmented AI: Search Tool Learning with LLM: Teaching LLMs to invoke search and make use of retrieved information
• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Abstract

Diffusion large language models (dLLMs) have recently drawn considerable attention within the research community as a promising alternative to autoregressive generation, offering parallel token prediction and lower inference latency. Yet, their parallel decoding potential remains largely underexplored, as existing open-source models still require nearly token-length decoding steps to ensure performance. To address this, we introduce dParallel, a simple and effective method that unlocks the inherent parallelism of dLLMs for fast sampling. We identify that the key bottleneck to parallel decoding arises from the sequential certainty convergence for masked tokens. Building on this insight, we introduce the core of our approach: certainty-forcing distillation, a novel training strategy that distills the model to follow its original sampling trajectories while enforcing it to achieve high certainty on masked tokens more rapidly and in parallel. Extensive experiments across various benchmarks demonstrate that our method can dramatically reduce the number of decoding steps while maintaining performance. When applied to the LLaDA-8B-Instruct model, dParallel reduces decoding steps from 256 to 30 on GSM8K, achieving an 8.5x speedup without performance degradation. On the MBPP benchmark, it cuts decoding steps from 256 to 24, resulting in a 10.5x speedup while maintaining accuracy. Our code is available at https://github.com/czg1225/dParallel

Problem

• Unlocking parallel decoding potential in diffusion language models
• Reducing sequential certainty convergence for masked tokens
• Maintaining performance while dramatically cutting decoding steps

Innovation

• Learnable parallel decoding for diffusion language models
• Certainty-forcing distillation accelerates token certainty
• Reduces decoding steps while maintaining model performance