Uno: Diffusion-Augmented LLM Uses LoRA Drafts and Ψ-Spec Sampler for Lossless Speedups

burkov · x · 2026-09-08

This paper targets a core LLM inefficiency: next-token prediction yields high-quality output but forces slow sequential generation, wasting accelerator capacity and inflating latency and cost — especially for long traces and concurrent agentic workloads.

Method — Uno (diffusion-augmented LLM): keep the full autoregressive weights trained with standard next-token prediction, then add lightweight LoRA adapters trained in a short Diffusion Distillation phase to propose blocks of tokens in parallel.

Inference: a new Ψ-Spec sampler draws parallel drafts from the diffusion pathway and verifies them against the frozen autoregressive distribution, aiming for lossless speedups — multiple tokens per step while preserving the exact output distribution of the strong base model.

Related event: Uno Uses Discrete Diffusion for Lossless LLM Speedups(3 posts)→

Original post →

More from Research

Research channel →