TL;DR: A team released ORL (On-Policy Residual Learning) plus a custom inference engine, claiming up to 6.9× throughput gains for diffusion LLMs at comparable quality, with all code and models open sourced.
Summary: Researchers published ORL, a training method that makes diffusion language models converge in fewer denoising passes by carrying forward the model's own prior predictions as residual context instead of re-predicting every token from scratch. Paired with their custom inference engine dlmrt, they report 2,788 tokens/s for LLaDA-2.1-mini on a single RTX 5090, 7,708 tokens/s for DiffusionGemma 26B MoE on 8× B200 in NVFP4, and 6.9× end-to-end speedup on QwenDLM-4B, plus 1.74× faster than DFlash speculative decoding at matched accuracy on a single H100 comparison. All figures are peak single-request (batch size 1) results on specific benchmarks rather than averages, and the models, training code, evaluation pipeline, and inference engine are open sourced.
Why it matters: Diffusion LLMs have been bottlenecked by the number of forward passes needed to converge, so a training-side fix that also pairs with an open inference engine is a credible path to production-grade parallel decoding. Worth reproducing the batch-size-1 benchmarks on your own hardware before planning around these numbers, since real-world and batched throughput will differ.
Source: reddit