Efficient-DLM: From Autoregressive to Diffusion Language Models, and Beyond in Speed

arXiv cs.CL / 5/1/2026

💬 OpinionIdeas & Deep AnalysisModels & Research

共有:

Key Points

Diffusion language models enable parallel, non-autoregressive generation, but they often learn less efficiently than autoregressive models when trained from scratch.
The paper proposes an AR-to-dLM conversion approach that targets limitations in existing methods’ attention patterns and training objectives to preserve task accuracy while improving speed.
It finds that keeping the pretrained AR weight distributions intact is crucial for effective conversion, and introduces a continuous pretraining scheme with a block-wise attention pattern that is causal across blocks yet bidirectional within each block.
To reduce a training–test mismatch in mask token distributions, the authors introduce a position-dependent token masking strategy that masks later tokens more during training.
Experiments yield the “Efficient-DLM” model family, where Efficient-DLM 8B reportedly achieves higher accuracy than prior AR/dLM baselines while delivering significantly higher throughput (e.g., 4.5x/2.7x faster than Dream 7B/Qwen3 4B).

Abstract

Diffusion language models (dLMs) have emerged as a promising paradigm that enables parallel, non-autoregressive generation, but their learning efficiency lags behind that of autoregressive (AR) language models when trained from scratch. To this end, we study AR-to-dLM conversion to transform pretrained AR models into efficient dLMs that excel in speed while preserving AR models' task accuracy. We achieve this by identifying limitations in the attention patterns and objectives of existing AR-to-dLM methods and then proposing principles and methodologies for more effective AR-to-dLM conversion. Specifically, we first systematically compare different attention patterns and find that maintaining pretrained AR weight distributions is critical for effective AR-to-dLM conversion. As such, we introduce a continuous pretraining scheme with a block-wise attention pattern, which remains causal across blocks while enabling bidirectional modeling within each block. We find that this approach can better preserve pretrained AR models' weight distributions than fully bidirectional modeling, in addition to its known benefit of enabling KV caching, and leads to a win-win in accuracy and efficiency. Second, to mitigate the training-test gap in mask token distributions (uniform vs. highly left-to-right), we propose a position-dependent token masking strategy that assigns higher masking probabilities to later tokens during training to better mimic test-time behavior. Leveraging this framework, we conduct extensive studies of dLMs' attention patterns, training dynamics, and other design choices, providing actionable insights into scalable AR-to-dLM conversion. These studies lead to the Efficient-DLM family, which outperforms state-of-the-art AR models and dLMs, e.g., our Efficient-DLM 8B achieves +5.4%/+2.7% higher accuracy with 4.5x/2.7x higher throughput compared to Dream 7B and Qwen3 4B, respectively.