New discrete-diffusion LLM speedup paper Uno called out for similarity to Orthrus

_akhaliq · x · 2026-09-09

A new Hugging Face paper page, "Unlocking Lossless Speedups in LLMs via Discrete Diffusion," presents Uno, which combines AR and diffusion generation while keeping the architecture (and its attention) unchanged for lossless speedups.

A commenter questioned whether the core framework is the same as Orthrus (arXiv:2605.12825), released four months earlier. Author Subham Sekhar Sahoo responded that Orthrus adds diffusion attention heads with bidirectional attention, changing the architecture, while Uno does not — acknowledging similarities but disputing that the core ideas match; a discussion will be added in a revision.

Related event: Uno: Discrete Diffusion Delivers Lossless LLM Speedups(4 posts)→

Original post →

More from Research

Research channel →