Diffusion adapters give lossless 3x LLM speedups with no draft model, Uno paper shows
kastnerkyle · x · 2026-09-06
A new paper introduces diffusion-augmented LLMs (Uno), which generate multiple tokens in parallel via diffusion while preserving the base autoregressive model's distribution.
How it works
- Parameters are decoupled: standard AR weights trained with next-token prediction, plus lightweight diffusion weights learned via a simple Diffusion Distillation phase that drafts tokens in parallel, verified exactly by the AR model.
- Unlike speculative decoding, no separate draft model is needed; unlike diffusion LLMs, quality of the underlying AR model is preserved.
- A family of samplers called Ψ-Spec enables lossless acceleration and inference-time scaling at fixed context length.
Results
- Uno models can be trained from scratch or built on existing open-weight AR LLMs.
- Higher throughput than leading speculative-decoding methods at every evaluated batch size, with up to 3x speedups over the base AR model.
- The 8B Uno outperforms the 26B open diffusion LM DiffusionGemma and the proprietary Mercury.
More from Research
- A real larva's connectome — 2,675 cells, 14,066 connections — now swims in a 3D simulation — eptwts · 2026-10-21
- Stanford's T2-z lets robots probe hidden physics with a poke, lifting success from 34% to 73% — stanfordnlp · 2026-10-21
- Fei-Fei Li team unveils FACT, a force-aware foundation model for robot dexterity — stanfordnlp · 2026-10-21
- Upcoming paper claims big models can handle latency-critical real-time robotics — ManlingLi_ · 2026-10-21
- NVIDIA open-sources Flora, a robotics data engine 125x cheaper than GPT-6 Astra — DrJimFan · 2026-10-21
- openRxiv explains its approach to AI amid growing researcher adoption — anshulkundaje · 2026-10-21