What they did

The authors separate each model into conventional autoregressive weights, trained by next-token prediction, and lightweight diffusion weights trained through an additional distillation phase. These diffusion components learn to propose multiple tokens simultaneously while the overall sampler targets the original autoregressive distribution.

They apply this approach both when training models from scratch and when augmenting existing open-weight models. Their Ψ-Spec sampler is evaluated against ordinary autoregressive decoding, speculative decoding, and diffusion language models on throughput and tasks involving tool use, coding, and long-context reasoning.

Key findings

• Uno delivered up to a 3× throughput improvement over its underlying autoregressive model.
• It exceeded the throughput of the evaluated speculative-decoding methods at every tested batch size, including the largest batch supported by the device, without requiring a separate draft model.
• The authors report that the sampler preserves the base model's autoregressive distribution rather than trading model quality for speed.
• An 8B-parameter Uno model outperformed the 26B DiffusionGemma and proprietary Mercury 2 across all evaluated benchmarks in agentic tool use, coding, and long-context reasoning.

Why it matters

The work offers a route around the sequential nature of next-token generation without replacing the familiar training objective or accepting an approximate target distribution. If the reported gains generalize, existing open-weight models could gain parallel decoding through a relatively small distillation step rather than a second draft model or complete retraining as a diffusion model.

Caveats

The abstract does not provide benchmark scores, hardware specifications, latency measurements, acceptance behavior, or the computational cost of diffusion distillation, so the breadth and practical economics of the speedups cannot be assessed here. The strongest quality comparison is also between models of different sizes and training histories; independent evaluation is needed to separate architectural benefits from data, training, and implementation differences.