arXiv AI

On Trajectory-Aware Training for Masked Diffusion Language Models

arXiv Computation and Language
Sep 23

Informed Masking: Structure-Aware Perturbation for Reinforcement Learning in Diffusion Large Language Models

Informed Masking (IM) is a new technique for aligning Diffusion Large Language Models (dLLMs) with Reinforcement Learning (RL). It identifies a systematic upstream/downstream token structure in dLLM rollouts and shows that masking downstream tokens creates better subproblems for likelihood estimation. When integrated into three state‑of‑the‑art dLLM RL methods on LLaDA‑8B‑Instruct, IM yields up to 2.01%, 8.68%, and 5.77% relative average gains on math and planning benchmarks while improving training stability.

By Xiaoyi Yu, Enver Sangineto, Pei Fu, Fiorenzo Parascandolo, Wenhui Tan, Ruikang Zhang, Rita Cucchiara, Ruihua Song, Jian Luan
arXiv Machine Learning
Oct 16

On-Policy Residual Learning for Diffusion Language Models

On-Policy Residual Learning for Diffusion Language Models (ORL) introduces a framework that learns residual context directly from model-generated trajectories, using an entropy-contraction objective to sharpen predictions. The approach co-designs algorithmic and system components—such as asynchronous residual capture, block-diffusion attention, and packed compute-balanced training—to enable practical scaling. Experiments on a 4B DLM converted from Qwen3-4B show up to 6.2× faster tokens per forward and 6.9× higher end-to-end throughput, while also improving speed-quality trade-offs over autoregressive models and transferring effectively to other DLMs like LLaDA-2.1-mini and DiffusionGemma.

By Yifan Yu, Yuqing Jian, Jisen Li, Yuezhou Hu, Tianyi Zhang, Xiaoxia Wu, Rui-Jie Zhu, Junxiong Wang, Zhongzhu Zhou, Fan Lai, James Zou, Chenfeng Xu
arXiv AI
Sep 3

Predict, Don't Iterate: Efficient Adaptive-Length Infilling for Diffusion Language Models

The paper introduces PILL, a new infilling technique for diffusion language models that eliminates the need for a preset initial length and reduces inference overhead. PILL uses probing-based length-free decoding, cutting down on extra forward passes and speeding up generation. Experiments across five diffusion models and eight benchmarks show PILL outperforms the strongest baseline with higher pass rates and BLEU-2 scores while running 1.82× faster.

By Haobo Xu, Sirui Chen, Yuanchen Bei, Lingjie Chen, Yuchen Yan, Dongqi Fu, Jingrui He, Hanghang Tong
arXiv Computation and Language
Oct 28

Self-Regularized Flow Language Models

The paper introduces Self-Regularized Flow Language Models (SERF), a self‑supervised framework that enhances continuous diffusion language models by leveraging the denoising trajectory. SERF employs noisy‑to‑clean self‑distillation, where an exponential moving average teacher provides additional supervision along the same interpolation path, and offers two variants—SERF‑L aligning intermediate representations and SERF‑P aligning output token distributions. Experiments on three continuous language‑model paradigms show significant gains in training speed, generation quality, and sampling efficiency, achieving up to 7.7× faster training, 24.3% MAUVE improvement, and 28.8% higher GSM8K accuracy. "whyItMatters":"The results demonstrate that noisy‑to‑clean self‑distillation can substantially improve continuous diffusion language models without external pretraining or inference changes, offering a practical path to faster, higher‑quality language generation."

By Theodoros Kouzelis, Spyros Gidaris, Nikos Komodakis
arXiv AI
Sep 18

Zarya: A Hybrid Autoregressive--Masked Diffusion Language Model with Flexible Training and Dual-Mode Inference

Zarya is a hybrid language model that jointly trains an autoregressive objective and a masked-diffusion objective within a single architecture. It structures training data into variable-size slots and uses a curriculum that gradually increases slot granularity, allowing a smooth transition from fine-grained AR learning to coarse-grained diffusion learning. At inference, Zarya offers two decoding modes—MDM sampling with first-hitting denoising and slotted speculative decoding that interleaves diffusion-based selection with autoregressive infilling—while fully decoupling training and inference regimes and supporting extensive configurability.

By Leonid Sinev, Ilya Koziev, Vladislav Leshchuk