ORL pushes diffusion LMs to a new speed-intelligence Pareto frontier with ~6.9x throughput gains
james_y_zou · x · 2026-10-16
Researchers from Together AI, UIUC, Stanford and others present ORL (On-Policy Residual Learning), answering why diffusion language models still matter alongside lossless speculative decoding: AR optimizes intelligence, spec decoding optimizes speed, and DLMs optimize the Pareto frontier of both.
Key points:
- ORL trains on fresh, verified residual trajectories from the current model and encourages entropy contraction for more decisive refinement
- Co-design across algorithms, training/inference engine, and kernels improves parallel decoding on QwenDLM, LLaDA-2.1-mini, and DiffusionGemma
- On QwenDLM: up to 6.2x higher tokens per forward, 6.9x end-to-end throughput on a single H100, and up to 5.2x faster training steps
- Matches quality while beating speculative decoding on speed, and wins even against lossy speculation
Live demos, code, and models are available; the paper is an arXiv preprint.
More from Research
- UCSD Researchers Propose Dataset Watermarking Detectable via Black-Box API Access — yuxiangw_cs · 2026-10-16
- EmbodiedEvo tops RoboDojo as agents inherit reusable skills and lessons — Aiden_Tech_Ai · 2026-10-16
- CHI reviewer laments: papers' Discussion sections now cite nothing — IanArawjo · 2026-10-16
- Casper: Understanding 90% of neurons ≠ understanding the network — StephenLCasper · 2026-10-16
- NVIDIA Launches cuPotential in CUDA-X to Bring AI Agents to Semiconductor Materials Simulation — tomaszbednarz · 2026-10-16
- Agentic Search Systems Are Rewarded Only for Final Answers, Raising Faithfulness Concerns — burkov · 2026-10-16