Unlocking Lossless Speedups in LLMs via Discrete Diffusion

TL;DR - Diffusion-augmented LLMs preserve an autoregressive model’s distribution while generating multiple tokens in parallel, enabling lossless inference acceleration. The proposed Uno models deliver up to 3× higher generation speed than their base models without a separate draft model.

• Uno separates standard next-token-trained autoregressive weights from lightweight diffusion weights learned through a low-overhead distillation phase.
• The Ψ-Spec sampler supports lossless acceleration and inference-time scaling at a fixed context length.
• Uno reportedly outperforms leading speculative-decoding methods in throughput across every evaluated batch size.
• The 8B model surpasses larger or proprietary diffusion LLMs on the evaluated agentic tool-use, coding, and long-context reasoning benchmarks.