Lexington Whalen

Yonggan Fu, Lexington Whalen, Zhifan Ye et al.

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.

Lexington Whalen, Nathan Bickel, Shash Comandur et al.

Literacy, or the ability to read and write, is a crucial indicator of success in life and greater society. It is estimated that 85% of people in juvenile delinquent systems cannot adequately read or write, that more than half of those with substance abuse issues have complications in reading or writing and that two-thirds of those who do not complete high school lack proper literacy skills. Furthermore, young children who do not possess reading skills matching grade level by the fourth grade are approximately 80% likely to not catch up at all. Many may believe that in a developed country such as the United States, literacy fails to be an issue; however, this is a dangerous misunderstanding. Globally an estimated 1.19 trillion dollars are lost every year due to issues in literacy; in the USA, the loss is an estimated 300 billion. To put it in more shocking terms, one in five American adults still fail to comprehend basic sentences. Making matters worse, the only tools available now to correct a lack of reading and writing ability are found in expensive tutoring or other programs that oftentimes fail to be able to reach the required audience. In this paper, our team puts forward a new way of teaching English spelling and word recognitions to grade school students in the United States: Wordification. Wordification is a web application designed to teach English literacy using principles of linguistics applied to the orthographic and phonological properties of words in a manner not fully utilized previously in any computer-based teaching application.

Lexington Whalen, Zhenbang Du, Haoran You et al.

Training diffusion models (DMs) requires substantial computational resources due to multiple forward and backward passes across numerous timesteps, motivating research into efficient training techniques. In this paper, we propose EB-Diff-Train, a new efficient DM training approach that is orthogonal to other methods of accelerating DM training, by investigating and leveraging Early-Bird (EB) tickets -- sparse subnetworks that manifest early in the training process and maintain high generation quality. We first investigate the existence of traditional EB tickets in DMs, enabling competitive generation quality without fully training a dense model. Then, we delve into the concept of diffusion-dedicated EB tickets, drawing on insights from varying importance of different timestep regions. These tickets adapt their sparsity levels according to the importance of corresponding timestep regions, allowing for aggressive sparsity during non-critical regions while conserving computational resources for crucial timestep regions. Building on this, we develop an efficient DM training technique that derives timestep-aware EB tickets, trains them in parallel, and combines them during inference for image generation. Extensive experiments validate the existence of both traditional and timestep-aware EB tickets, as well as the effectiveness of our proposed EB-Diff-Train method. This approach can significantly reduce training time both spatially and temporally -- achieving 2.9$\times$ to 5.8$\times$ speedups over training unpruned dense models, and up to 10.3$\times$ faster training compared to standard train-prune-finetune pipelines -- without compromising generative quality. Our code is available at https://github.com/GATECH-EIC/Early-Bird-Diffusion.