跳转至

文章背景与核心概要

扩散大语言模型（dLLMs）通过在单次前向传播中同时预测多个Token（或掩码），实现了更快的语言生成速度。然而，在激进的并行解码策略下，这些模型往往在早期的去噪阶段产生不可靠的预测，从而在后续阶段引发级联错误。

为了克服这一局限性，作者引入了一致性强迫（Consistency Forcing, CForce），这是一种新颖的蒸馏方法，旨在将早期掩码预测与后期更精致的预测进行对齐。通过在预先收集的自展开轨迹（self-rollout trajectories）上训练模型，CForce显著改善了训练与推理的对齐性。核心亮点包括： * 置信度自适应KL散度： 一种结合了前向和反向Kullback-Leibler（KL）散度优势的定制蒸馏目标。 * 理论基础： 解释CForce如何最小化早期预测误差的数学保证。 * 普适性： 该框架无缝扩展至掩码到Token解码以及支持编辑的解码（其中后期的Token到Token精炼监督早期的掩码状态）。 * 性能提升： 在LLaDA模型上的实验表明，它增强了速度与质量的权衡，特别是在高并行度方案下。

CForce: Boosting Parallel Decoding for dLLMs via Consistency Forcing

CForce: Boosting Parallel Decoding for dLLMs via Consistency Forcing

arXiv ID: arXiv:2608.13925
Subjects: Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Computation and Language (cs.CL)
Authors: Yuji Ren, Chenkai Xu, Zhuocheng Gong, Jianguo Li, Zhijie Deng
Submitted: 14 August 2026
License: Creative Commons Attribution 4.0 International (View License Icon: license icon)

arXiv ID: arXiv:2608.13925
Subjects: Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Computation and Language (cs.CL)
Authors: Yuji Ren, Chenkai Xu, Zhuocheng Gong, Jianguo Li, Zhijie Deng
Submitted: 14 August 2026
License: Creative Commons Attribution 4.0 International (View License Icon: license icon)

📌 Summary

📌 Summary

Diffusion Large Language Models (dLLMs) achieve faster language generation by predicting multiple tokens (or masks) simultaneously in a single forward pass. However, under aggressive parallel decoding strategies, these models often produce unreliable predictions during early denoising stages, causing cascading errors in later stages.

Diffusion Large Language Models (dLLMs) achieve faster language generation by predicting multiple tokens (or masks) simultaneously in a single forward pass. However, under aggressive parallel decoding strategies, these models often produce unreliable predictions during early denoising stages, causing cascading errors in later stages.

To overcome this limitation, the authors introduce Consistency Forcing (CForce), a novel distillation method designed to align early-stage mask predictions with those of later, more refined stages. By training the model on pre-collected self-rollout trajectories, CForce significantly improves training-inference alignment. Key highlights include: * Confidence Adaptive KL Divergence: A custom distillation objective that combining the strengths of both forward and reverse Kullback-Leibler (KL) divergence. * Theoretical Foundations: Mathematical guarantees explaining how CForce minimizes early-stage prediction error. * Universal Applicability: The framework seamlessly extends to both mask-to-token decoding and edit-capable decoding (where later token-to-token refinements supervise earlier masked states). * Improved Performance: Experiments on LLaDA models show enhanced speed-quality trade-offs, particularly under high-parallelism regimes.

To overcome this limitation, the authors introduce Consistency Forcing (CForce), a novel distillation method designed to align early-stage mask predictions with those of later, more refined stages. By training the model on pre-collected self-rollout trajectories, CForce significantly improves training-inference alignment. Key highlights include: * Confidence Adaptive KL Divergence: A custom distillation objective that combines the strengths of both forward and reverse Kullback-Leibler (KL) divergence. * Theoretical Foundations: Mathematical guarantees explaining how CForce minimizes early-stage prediction error. * Universal Applicability: The framework seamlessly extends to both mask-to-token decoding and edit-capable decoding (where later token-to-token refinements supervise earlier masked states). * Improved Performance: Experiments on LLaDA models show enhanced speed-quality trade-offs, particularly under high-parallelism regimes.