SAS Turns Decoding Order Into Trainable Inference Strategy

furongh · x · 2026-07-06

SAS (Self-Aware Scheduler) transforms the decoding order into a trainable inference strategy. Instead of training the entire language model, it trains a lightweight scheduler while keeping the diffusion LM frozen. At each step, the scheduler selects the next masked position to reveal, turning the scheduling process itself into a "thinking process."

Related event: ICML 2026 Paper SAS: Optimizing Thought Scheduling in Diffusion LMs(15 posts)→

Original post →

More from Research

Research channel →