dUltra: Ultra-Fast Diffusion Language Models via Reinforcement Learning

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Chen, Shirui, Jiao, Jiantao, Ratliff, Lillian J., Zhu, Banghua
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866908817120296960
author Chen, Shirui
Jiao, Jiantao
Ratliff, Lillian J.
Zhu, Banghua
author_facet Chen, Shirui
Jiao, Jiantao
Ratliff, Lillian J.
Zhu, Banghua
contents Masked diffusion language models (MDLMs) offer the potential for parallel token generation, but most open-source MDLMs decode fewer than 5 tokens per model forward pass even with sophisticated sampling strategies, limiting their parallel generation potential. Existing acceleration methods either rely on fixed confidence-based heuristics or use distillation-based approaches that finetune MDLMs on trajectories generated by a base model, which can become off-policy during finetuning and restrict performance to the quality of the base model's samples. We propose \texttt{dUltra}, an on-policy reinforcement learning framework based on Group Relative Policy Optimization (GRPO) that learns unmasking strategies for efficient parallel decoding. dUltra introduces an unmasking planner head that predicts per-token unmasking likelihoods under independent Bernoulli distributions. We jointly optimize the base diffusion LLM and the unmasking order planner using reward signals combining verifiable reward, distillation reward, and the number of unmasking steps. Across mathematical reasoning and code generation tasks, dUltra achieves superior accuracy-efficiency trade-offs compared to state-of-the-art heuristic (Fast-dLLM) and distillation baselines (d3LLM, dParallel), demonstrating that learned unmasking trajectories through on-policy RL enable better exploitation of parallel generation in MDLMs. Code and checkpoints are released at https://github.com/chinsengi/dUltra-os.
format Preprint
id arxiv_https___arxiv_org_abs_2512_21446
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle dUltra: Ultra-Fast Diffusion Language Models via Reinforcement Learning
Chen, Shirui
Jiao, Jiantao
Ratliff, Lillian J.
Zhu, Banghua
Machine Learning
Artificial Intelligence
Masked diffusion language models (MDLMs) offer the potential for parallel token generation, but most open-source MDLMs decode fewer than 5 tokens per model forward pass even with sophisticated sampling strategies, limiting their parallel generation potential. Existing acceleration methods either rely on fixed confidence-based heuristics or use distillation-based approaches that finetune MDLMs on trajectories generated by a base model, which can become off-policy during finetuning and restrict performance to the quality of the base model's samples. We propose \texttt{dUltra}, an on-policy reinforcement learning framework based on Group Relative Policy Optimization (GRPO) that learns unmasking strategies for efficient parallel decoding. dUltra introduces an unmasking planner head that predicts per-token unmasking likelihoods under independent Bernoulli distributions. We jointly optimize the base diffusion LLM and the unmasking order planner using reward signals combining verifiable reward, distillation reward, and the number of unmasking steps. Across mathematical reasoning and code generation tasks, dUltra achieves superior accuracy-efficiency trade-offs compared to state-of-the-art heuristic (Fast-dLLM) and distillation baselines (d3LLM, dParallel), demonstrating that learned unmasking trajectories through on-policy RL enable better exploitation of parallel generation in MDLMs. Code and checkpoints are released at https://github.com/chinsengi/dUltra-os.
title dUltra: Ultra-Fast Diffusion Language Models via Reinforcement Learning
topic Machine Learning
Artificial Intelligence
url https://arxiv.org/abs/2512.21446