PaperlayerView exact revision on arXiv

Source-linked AI summary

dParallel: Learnable Parallel Decoding for dLLMs

Zigeng Chen, Gongfan Fang, Xinyin Ma, Ruonan Yu, Xinchao Wang

arXiv:2509.26488v1cs.CL

TL;DR

Open-source dLLMs retain parallel-token prediction but still need nearly sequence-length decoding, motivating faster sampling without quality loss. dParallel addresses this with certainty-forcing distillation, and experiments report large step reductions with maintained performance. The method’s gains remain dependent on the pretrained model and the limited training-data scale leaves further performance and generalization unexplored.

  • Problem

    Open-source dLLMs still require nearly sequence-length decoding steps to preserve generation quality, leaving their parallel-decoding potential underexplored.

  • Method

    Certainty-forcing distillation preserves the original sampling trajectory while minimizing predictive entropy so masked tokens reach high certainty more rapidly and in parallel.

  • Results

    Across benchmarks, dParallel substantially reduces decoding steps while maintaining comparable performance, including 88% fewer steps on GSM8K and 91% fewer on MBPP.

  • Takeaways & Limitations

    dParallel establishes a learnable baseline for parallel decoding and provides a foundation for future few-step and parallel dLLM research.

  • Takeaways & Limitations

    Performance gains depend on the pretrained dLLM, and the current training uses only around 10k math problems, limiting demonstrated scope for improving performance and generalization.

Abstract

from arXiv · show

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

1 INTRODUCTION

dParallel targets the sequential certainty convergence that limits parallel decoding in open-source dLLMs. Its certainty-forcing distillation preserves sampling trajectories while encouraging masked tokens to become confident more rapidly and in parallel, reducing decoding steps without sacrificing performance.

  • Open-source dLLMs still require sequence-length-proportional decoding steps to maintain generation quality, limiting inference efficiency.
  • Although masked tokens are predicted in parallel, their certainty converges left to right, preventing reliable simultaneous token determination.
  • Certainty-forcing distillation self-distills along the original semi-autoregressive trajectory while minimizing entropy on correctly predicted masked tokens.
  • 88% fewer decoding steps on GSM8K yields an 8.5× speedup for LLaDA-8B-Instruct without sacrificing accuracy.
  • 91% fewer decoding steps on MBPP delivers a 10.5× acceleration while maintaining performance.
  • The method is presented as a learnable approach that establishes a baseline for parallel decoding and supports future few-step dLLM research.

2 RELATED WORKS

Related work positions dLLMs as masked-token generators with parallel-decoding potential and highlights ongoing efforts to accelerate their inference. These efforts target either per-step cost or the number of required decoding steps.

  • dLLMs model language in discrete space by predicting masked tokens, offering potential for fast and parallel decoding.
  • LLaDA and Dream demonstrate performance comparable to autoregressive LLMs at the billion-parameter scale.
  • Caching mechanisms and token dropping reduce the time cost of individual dLLM decoding steps.
  • Other approaches accelerate inference by optimizing parallel sampling algorithms to reduce the required number of decoding steps.

3 PRELIMINARIES

Masked diffusion language models generate text through forward masking and reverse denoising rather than strict left-to-right prediction. Their sampling process jointly predicts masked tokens and dynamically remasks positions for further refinement.

  • MDLMs formulate generation as forward masking corruption followed by reverse denoising recovery.
  • The reverse process uses a mask predictor to recover the clean sequence from a noisy input.
  • At each reverse step, the model predicts all masked tokens jointly.
  • Training uses negative log-likelihood restricted to masked positions, which upper bounds the model’s negative log-likelihood.
  • Sampling predicts distributions for masked tokens, samples provisional tokens, and dynamically remasks positions needing further refinement.

4 METHOD

The method analyzes why dLLM certainty converges sequentially despite parallel masked-token prediction, then introduces certainty-forcing distillation to preserve generation trajectories while accelerating parallel certainty convergence.

  • 4.1 THE BARRIERS TO PARALLEL DECODING: High token confidence correlates positively with generation correctness, making certainty a useful training signal for decoding quality.Higher-confidence resolutions achieve higher accuracy, while low-confidence commitments frequently produce errors.
  • 4.1 THE BARRIERS TO PARALLEL DECODING: dLLMs theoretically predict masked tokens in parallel, but their certainty converges sequentially through iterative context enrichment.Only a few tokens near known context typically become highly confident at each step, leaving most tokens low-confidence until new context is available.
  • 4.1 THE BARRIERS TO PARALLEL DECODING: Sequential certainty convergence is the fundamental bottleneck because early multi-token commitments can introduce low-confidence predictions, cascading errors, and performance degradation.Reliable parallel decoding therefore requires many token positions to reach high certainty simultaneously.
  • 4.2 CERTAINTY-FORCING DISTILLATION: Certainty-forcing distillation self-distills a dLLM along its original generation trajectory while directly training predictive certainty toward faster parallel convergence.The strategy combines trajectory consistency with entropy minimization over correctly predicted masked tokens.
  • 4.2 CERTAINTY-FORCING DISTILLATION: Training simulates semi-autoregressive generation by masking an active block while preserving earlier context and fully masking future blocks.The student is trained on the active block to replicate the target trajectory and achieve maximal certainty in parallel.
  • 4.2 CERTAINTY-FORCING DISTILLATION: The objective combines consistency loss with certainty-forcing loss, whose entropy term sharpens predictions for correctly predicted masked tokens.The balancing coefficient β controls the trade-off between matching the teacher trajectory and enforcing high certainty.

5 EXPERIMENTS

Experiments evaluate dParallel across open-source dLLMs and benchmarks, showing faster decoding while preserving performance. Ablations and confidence analyses attribute the gains to certainty-forcing distillation and its joint consistency–certainty objectives.

  • Experimental Setup: Experiments evaluate LLaDA and Dream across mathematics and code-generation benchmarks using accuracy, decoding steps, latency, and speedup.The baselines include official inference, few-step decoding, confidence-threshold decoding, and caching-based acceleration.
  • Main Results: More than 8 tokens decoded per step enables LLaDA to maintain performance while substantially increasing parallelism.Confidence-threshold remasking preserves accuracy but averages only 3–4 tokens decoded per step.
  • Main Results: At 9.4× speedup on LLaDA GSM8K, dParallel achieves 16.5% higher accuracy than confidence-threshold decoding.On HumanEval, it improves accuracy by 21.3% at the same 9.3× speedup; Dream shows a similar trade-off curve.
  • Certainty Convergence: Certainty-forcing reshapes sequential token-certainty convergence into a faster, more parallel process.The analysis compares average token confidence at the 8th and 16th decoding steps for LLaDA-8B-Instruct on GSM8K.
  • Ablation Study: Removing the certainty loss leaves sequential convergence largely unchanged, while removing consistency produces high speed but sharply lower performance.The full strategy combines trajectory consistency with certainty optimization.
  • Ablation Study: A fixed 50% masking ratio provides the best balance of acceleration and accuracy, whereas higher, lower, and random ratios degrade accuracy.This masking-ratio comparison trains LLaDA for one epoch.

6 CONCLUSION

The conclusion presents dParallel as a learnable approach for reducing dLLM decoding steps through certainty-forcing distillation. The method maintains trajectory consistency while encouraging faster, more parallel high-certainty predictions.

  • 6 CONCLUSION: dParallel reduces dLLM decoding steps by combining trajectory consistency with high-certainty masked-token prediction.The paper positions this strategy as a new training paradigm and baseline for parallel decoding.

A MORE IMPLEMENTATION DETAILS

Implementation details describe training configurations across three base models and report a memory-efficient setup. Additional results show substantial decoding-step reductions for LLaDA-1.5 while preserving accuracy and reliability.

  • Training Configuration: Training configurations for certainty-forcing distillation are provided across three base models.LLaDA data use fixed 384-token sequences, while Dream responses retain their original 256-token length.
  • Training Configuration: 23 GB of GPU memory is sufficient for training with PEFT and shorter sequence lengths.Training uses two NVIDIA H100 GPUs with an effective global batch size of 64.

B MORE EXPERIMENTAL RESULTS

Additional analysis shows that dParallel accelerates certainty convergence across token positions. Applied to LLaDA-1.5, the method substantially compresses decoding while preserving task performance.

  • Additional Experimental Results: LLaDA-1.5 requires only 26–46 decoding steps instead of the baseline’s 256 across four benchmarks.The reduction substantially accelerates generation while preserving accuracy and reliability.
  • Certainty Analysis: Original LLaDA certainty converges sequentially, whereas dParallel produces faster and more parallel convergence across the first 160 positions.The confidence analysis covers the initial 16 decoding steps on GSM8K.

C CASE STUDY

The case studies show that dParallel substantially reduces decoding steps while maintaining generation quality across multiple tasks.

  • dParallel significantly reduces decoding steps in the presented case studies.The paper reports this pattern across Figures 7, 8, and 9.
  • The reduced-step decoding maintains generation quality in the case studies.

D LIMITATIONS AND FUTURE WORK

The method’s efficiency gains depend on the pretrained dLLM’s performance, while future work targets broader training and responsible deployment. The study uses public data and does not involve human subjects or sensitive personal data.

  • Limitations: The method cannot significantly improve performance when the pretrained dLLM is weak.Its inference-efficiency gains and strong accuracy remain dependent on the base model’s performance.
  • Future Work: Future work will extend certainty-forcing to pretraining and substantially scale up the training data.The authors currently use a relatively small dataset of around 10k math problems.
  • Ethics: The study uses publicly available licensed datasets and excludes human subjects and sensitive personal data.The authors also acknowledge potential misuse risks and encourage responsible use.

F REPRODUCIBILITY STATEMENT

The work emphasizes reproducibility through detailed documentation, public datasets, preprocessing descriptions, pseudocode, implementation details, and released source code.

  • Model architectures, training objectives, and experimental settings are described in the main text and Appendix.
  • The study documents preprocessing steps for its publicly available datasets.
  • Pseudocode, implementation details, and source code are provided to facilitate replication.
Loading 2509.26488v1…