Scaling up Masked Diffusion Models on Text

Masked diffusion models (MDMs) have shown promise in language modeling, yet their scalability and effectiveness in core language tasks, such as text generation and language understanding, remain underexplored. This paper establishes the first scaling law for MDMs, demonstrating a scaling rate comparable to autoregressive models (ARMs) and a relatively small compute gap. Motivated by their scalability, we train a family of MDMs with up to 1.1 billion (B) parameters to systematically evaluate their performance against ARMs of comparable or larger sizes. Fully leveraging the probabilistic formulation of MDMs, we propose a simple yet effective \emph{unsupervised classifier-free guidance} that effectively exploits large-scale unpaired data, boosting performance for conditional inference. In language understanding, a 1.1B MDM shows competitive results, outperforming the larger 1.5B GPT-2 model on four out of eight zero-shot benchmarks. In text generation, MDMs provide a flexible trade-off compared to ARMs utilizing KV-cache: MDMs match the performance of ARMs while being 1.4 times faster, or achieve higher quality than ARMs at a higher computational cost. Moreover, MDMs address challenging tasks for ARMs by effectively handling bidirectional reasoning and adapting to temporal shifts in data. Notably, a 1.1B MDM breaks the \emph{reverse curse} encountered by much larger ARMs with significantly more data and computation, such as Llama-2 (13B) and GPT-3 (175B). Our code is available at \url{https://github.com/ML-GSAI/SMDM}.