Unlocking Lossless Speedups in LLMs via Discrete Diffusion

Thanks for brining this to our attention.

The core idea of our method Uno is to achieve AR and diffusion generation keeping the architecture (with its attention) unchanged, but the paper that you shared changes the architecture by adding diffusion attention heads.

Not only that, it uses bidirectional attention for the diffusion blocks.

There are a lot of similarities, yes, but i respectfully disagree with this characterization that the core ideas are the same.

We’ll add a discussion in the next revision

Diffusion LLMs have two major issues wrt AR LLMs:

1. Poor quality,
2. Slow inference at high batch sizes.

To overcome these limitations, we introduce diffusion-augmented LLMs, a new class
of models that defines an AR model distribution while using diffusion to draw multiple
tokens in parallel from that distribution. We decouple the parameters of these models into
two sets:

• AR weights: Trained using next-token prediction loss to define an AR distribution.
• Diffusion weights: Trained to generate multiple tokens in parallel from the AR distribution.
  under a single unified LLM architecture.

We name our model Uno. To sample from it, we propose $\Psi$-Spec sampler, which enables provably lossless multi-token prediction from the AR distribution.

• Unlike speculative decoding methods, Uno does not require a separately trained draft model and achieves higher throughput across all batch sizes.
• Unlike self-speculative decoding approaches, Uno is lossless.

Results

Our model Uno

• Achieves higher throughput than speculative-decoding methods at every batch size, and
• Delivers up to 3× speedups over the base AR model, including at the largest batch size supported by the
  device.

Notably, our 8B Uno model outperforms the leading open d-LLM, the 26B DiffusionGemma, and the proprietary Mercury 2 across all evaluated benchmarks in agentic tool use, coding, and long-context reasoning.

Lossless is doing a lot of work in that title. Tokens per second on a clean benchmark doesn't tell me what happens at batch 32 on an H100 with KV cache pressure — that's where AR models are actually fast. The real question is denoising steps: how many passes does discrete diffusion need to match AR quality? If it's more than two or three, the parallel gain gets eaten by sampling overhead and you're back to trading latency for throughput. I'd want the wall-clock number at a realistic batch size before I believe "lossless speedup" survives contact with prod.

This is an automated message from the Librarian Bot. I found the following papers similar to this paper.

The following papers were recommended by the Semantic Scholar API

• Retrofitting Linear Attention into Diffusion Language Models (2026)
• AngelSpec: Towards Real-World High Performance Inference with Speculative Decoding (2026)
• AdaMTP: An Adaptive Training Paradigm for Multi-Token Prediction (2026)
• Trajectory-Level Speculative Decoding for Diffusion Language Models (2026)
• Rethinking the Generation Order of Block Diffusion Language Models (2026)
• Accepted Prefixes Are Not All You Need: A Negative Result on PEFT-Based Block-Diffusion Drafting (2026)
• DiffusionGemma Technical Report (2026)

Please give a thumbs up to this comment if you found it helpful!

If you want recommendations for any Paper on Hugging Face checkout this Space

You can directly ask Librarian Bot for paper recommendations by tagging it in a comment: @librarian-bot recommend

Nice work, congratulations! I know it's hard to compete with the autoregressive baseline models, given how much longer AR inference has been optimized, so this is a really exciting result.

I was a little surprised, though, not to see our earlier LAD: LoRA-Adapted Diffusion work mentioned in the related work. In LAD (EMNLP 2025), we similarly showed that a pretrained AR LLM can be adapted for parallel diffusion-based generation using only lightweight LoRA adapters, while keeping the pretrained weights frozen: https://aclanthology.org/2025.emnlp-demos.8/

Of course, Uno goes considerably further, adding a lot of new features: AR-distribution distillation, retaining causal attention (which I kind of regret, looking at the promised properties of diffusion models!), and the Ψ-Spec verification mechanism.

But the underlying idea of augmenting an existing AR model with only LoRA parameters trained specifically for diffusion generation seems quite closely related to LAD. I actually mentioned LAD on a different paper of your group here on Hugging Face back in March, (dLLM: https://huggingface.co/papers/2602.22661), so I had hoped it might make it into the discussion of prior approaches.

In any case, really nice work, and very cool to see this general direction pushed so much further!