Researchers introduced ORL (On-Policy Residual Learning), a technique that accelerates diffusion language models while preserving quality. Using ORL and a custom inference engine (dlmrt), they achieved 2,788 tokens/s on a single RTX 5090 with LLaDA‑2.1‑mini and up to 7,708 tokens/s on eight B200 GPUs with DiffusionGemma 26B MoE, delivering up to 6.9× higher throughput on QwenDLM‑4B.