Learning When to Commit: Schedule Distillation for Fast Diffusion Language Models
AuthorsAidar Myrzakhan, Zhiqiang Shen
AffiliationsVILA Lab, MBZUAI
dSpark teaches diffusion language models which tokens to reveal first, aiming to generate answers much faster without sacrificing quality.
Key results
An uncertain reveal made 5.1× more other positions correct than a confident reveal.
An uncertain reveal reduced other positions’ surprisal 4.7× more than a confident reveal.
dSpark achieved 10.12× the base model’s decoding parallelism.
dSpark achieved 5.12× the base model’s decoding parallelism.
Tokens per second on an NVIDIA A100; reported as 16.6× the base model’s throughput.
What the paper found
dSpark accelerates diffusion language models by learning when to reveal tokens, rather than simply committing whichever predictions are most confident. Its offline schedule construction promotes plausible tokens when revealing them reduces uncertainty across other masked positions; on GSM8K, revealing an uncertain token made 5.1× more other positions correct and reduced their surprisal 4.7× more than revealing a confident token. The team distills these schedules into LLaDA-8B-Instruct and Dream-7B-Instruct, using a KL-divergence penalty on failed trajectories to limit accuracy loss. Across GSM8K, MATH, HumanEval, and MBPP, dSpark reaches 10.12× the base model’s decoding parallelism on LLaDA and 5.12× on Dream for GSM8K, while LLaDA accuracy is 78.9%. On an NVIDIA A100, LLaDA generates 249 tokens per second on GSM8K, a reported 16.6× throughput gain over the base model. The central result is that useful context—not confidence alone—can guide faster parallel decoding while preserving task performance.
Original abstract
Diffusion large language models (dLLMs) can generate multiple tokens in parallel, but often require many decoding rounds to complete a response. Efficient generation depends not only on which tokens are easy to predict, but also on how committing to them helps predict the remaining positions. A highly predictable token may provide little useful context, while revealing a less predictable token can help several other positions become predictable sooner. We propose dSpark, a new schedule distillation-based framework for learning when to commit tokens for faster diffusion language generation. To achieve this, we construct unmasking schedules that prioritize tokens whose reveal makes other positions easier to predict, reducing uncertainty earlier in decoding. We then distill these schedules into the model to enable faster parallel sampling, with a KL-divergence regularizer to help preserve generation quality. We extensively evaluate our approach across mathematical reasoning and code generation tasks. On GSM8K, our method achieves remarkable 10.12$\times$ and 5.12$\times$ speedups over the various baselines on LLaDA-8B-Instruct and Dream-7B-Instruct, respectively, while maintaining superior task performance. Our code is available at https://github.com/VILA-Lab/dSpark.
Read the original paperMore in Diffusion Models
Browse all 82 papers →FewStepRL: Fine-tuning Few-step Diffusion Models with Reinforcement Learning
Haitao Lin, Minsi Ren, Yang Li, Xingtong Ge, Zheyuan Hu, Peiyan Hu, Tailin Wu
FewStepRL adapts multiple reinforcement-learning fine-tuning methods to make reward-driven improvements to fast, few-step diffusion models.
JVT: Just Video Transformer makes pixel-space video diffusion simple
Jianjun Zhou, Hua Geng, Haoyu Guo
JVT shows that a straightforward Transformer can generate video directly in pixels—and even produce matching depth-like geometry alongside it.
Restoring Diversity in Distilled Image and Video Generators by Probing Activations
Furkan Horoz, Ozgur Kara, Kaan Oktay, Adil Kaan Akan, Atakan Kara, Anand Bhattad, James Matthew Rehg
The paper restores variety in fast distilled image and video generators by finding and correcting a collapse in their internal activation directions.