They make language models generate multiple words at once without losing accuracy.

Large Language Models, or LLMs, usually create text one piece at a time. This slow, sequential step is called autoregressive generation. Because each word depends on the one right before it, making long texts takes a lot of time. To solve this bottleneck, researchers created Uno. Uno splits its setup into two parts: standard weights that predict the next token, and lightweight diffusion weights that draw multiple tokens in parallel from that same distribution. The diffusion weights are trained with a quick step called Diffusion Distillation, which adds negligible overhead to standard training. Uno uses a sampler called $Ψ$-Spec to speed up generation without losing any output quality. Unlike older shortcuts, it does not need a separate draft model. Uno achieves up to 3× speedups over base models, beats top speculative-decoding methods in throughput, and beats larger systems in coding, tool use, and long reasoning.

Why it matters

This means language models can work much faster, saving time and energy for everyone from students to companies using AI tools. Faster generation allows for more responsive applications and real-time interactions.

Questions answered