SAS Teaches Diffusion LMs Self-Aware Scheduling

An ICML 2026 paper introduces SAS, a self-aware scheduling method for diffusion language models that rewards unmasking orders using the model’s own frozen path likelihoods. It needs no human-written reasoning traces and boosts Sudoku accuracy from 82.0% to 91.8%, or 97.5% with a second fine-tuning stage, while also improving GSM8K from 64% to 76% and MBPP from 39.5% to 41%.

2026-07-06 ~ 2026-07-07 · 15 related posts