Unlocking Lossless Speedups in LLMs via Discrete Diffusion
Uno uses lightweight diffusion layers to sample several tokens in parallel while preserving the base autoregressive model’s distribution.
The paper says its diffusion-augmented LLMs keep standard next-token-trained AR weights and add a small diffusion distillation stage for multi-token generation. Its Ψ-Spec samplers are presented as lossless acceleration without a separate draft model. The authors report up to 3× speedups over the base AR model and higher throughput than leading speculative-decoding methods at every evaluated batch size. They also say an 8B Uno beats DiffusionGemma 26B and Mercury 2 on their evaluated agentic tool use, coding, and long-context reasoning benchmarks. HF Daily Papers' note
The paper says its diffusion-augmented LLMs keep standard next-token-trained AR weights and add a small diffusion distillation stage for multi-token generation. Its Ψ-Spec samplers are presented as lossless acceleration without a separate draft model. The authors report up to 3× speedups over the base AR model and higher throughput than leading speculative-decoding methods at every evaluated batch size. They also say an 8B Uno beats DiffusionGemma 26B and Mercury 2 on their evaluated agentic tool use, coding, and long-context reasoning benchmarks. HF Daily Papers' note
score 6