Scheduling Thoughts: Learning the Order of Thought in Diffusion Language Models

The paper introduces Self-Aware Scheduling (SAS) for masked diffusion language models, which optimizes the order of token unmasking to enhance generation quality. By deriving a tractable upper bound on decoding mismatch, SAS employs Group Relative Policy Optimization to learn a lightweight order policy, resulting in significant performance improvements: from 82.0% to 91.8% accuracy in Sudoku puzzles using a 1B MDM model, and increased pass rates on GSM8K and MBPP benchmarks with the 8B LLaDA model. This approach provides a principled method for order selection, benefiting practitioners by improving model performance across various tasks.