Researchers led by Yifan Yu introduced ORL, an on‑policy residual learning method for diffusion language models. ORL trains the model on its own prediction paths and adds an entropy‑contraction loss to sharpen outputs. On a 4 B model converted from Qwen3‑4B, ORL boosted tokens per forward by up to 6.2× and overall throughput by about 6.9×, matching or beating quality of autoregressive specs. The same technique gave a 1.5× speed gain on LLaDA‑2.1‑mini and modest gains on DiffusionGemma.
Why it matters
Developers can deploy faster, high‑quality text generation on consumer‑grade GPUs, reducing cost and latency for applications like chatbots.
