Diffusion-augmented LLM Uno delivers lossless 2.2x speedup over autoregressive generation
HongyiWang10 · x · 2026-09-18
IFMAI introduces Uno, a diffusion-augmented LLM aimed at the sequential token-by-token inference bottleneck of autoregressive models.
Key ideas:
- Parameters are split into standard NTP-trained AR weights plus lightweight diffusion weights that draw multiple tokens in parallel
- The diffusion weights are learned via a simple Diffusion Distillation phase with negligible overhead to existing training pipelines
- A family of samplers, Ψ-Spec, enables lossless acceleration and inference-time scaling at fixed context length — no draft model (unlike speculative decoding) and no quality loss (unlike diffusion LLMs)
Results: Uno can be trained from scratch or bolted onto open-weight AR LLMs; K2-Horizon-7B beats SOTA diffusion methods on both quality and throughput with up to 2.2x speedup, and outperforms leading speculative-decoding throughput at every evaluated batch size. Paper on arXiv (2609.04010), model released.
More from Research
- ETH's anthropomorphic robot hand learns to walk on its own fingers, untethered — ethz · 2026-09-18
- Conjectures claims two Erdős problem solutions per day via Bittensor, including a 55-year-old open question — markjeffrey · 2026-09-18
- New preprint: grammaticality emerges as an organizing principle in LM representations — TuhinChakr · 2026-09-18
- NVIDIA tutorial: memory-driven self-model agent hits 90.9% vs 82.8% RAG baseline — dl_weekly · 2026-09-18
- New paper: Tri-Metric Router cuts long-context RAG OOM failures to 0% on a T4 — chaumian · 2026-09-18
- Data Lab Argues Truly General Synthetic Data Matters More Than Architecture — Paimaamu · 2026-09-18