Omer Lux

[Omer Lux @OmerLux]

I'm proud to share that our paper, "Plan for Speed: Dilated Scheduling for Masked Diffusion Language Models", has been accepted to #ICML2026! 🎉

Huge thanks to Haim Permuter and @NachmaniEliya for the great collaboration.

[Omer Lux @OmerLux]

🧵1/ Meet Plan for Speed: Dilated Unmasking Scheduler (DUS) for Masked Diffusion LMs - a drop-in, inference-only planner that shatters the speed–quality trade-off.

📄 Paper: https://t.co/qYZee0ZL37
🌐 Site and Demos: https://t.co/NhDeKO5vPV

[Omer Lux @OmerLux]

🧵6/ Results:
Preserves ≳25% quality at 5–10× fewer passes for math, code, and general knowledge reasoning benchmarks, revealing MDLMs’ true parallel potential.

Huge thanks to Haim Permuter and @NachmaniEliya for their collaboration!

#DiffusionLM #LLM #NLP #NonAutoregressive

[Omer Lux @OmerLux]

🧵1/ Meet Plan for Speed: Dilated Unmasking Scheduler (DUS) for Masked Diffusion LMs - a drop-in, inference-only planner that shatters the speed–quality trade-off.

📄 Paper: https://t.co/qYZee0ZL37
🌐 Site and Demos: https://t.co/NhDeKO5vPV

[Omer Lux @OmerLux]

🧵5/ Why it works:
Under a 1st-order Markov view, distant tokens share almost no info, so DUS’s grouping slashes joint entropy near-optimally - far better than self-confidence schedulers.