Improving Discrete Diffusion Unmasking Policies Beyond Explicit Reference Policies

Researchers developed a learned scheduler for masked diffusion models (MDMs) in language modeling that outperforms traditional rule-based approaches. The new method uses a KL-regularized Markov decision process framework and demonstrated significant improvements, including 20.1% gains over random scheduling and 11.2% over max-confidence approaches on benchmark tests.

• →Masked diffusion models for language generation are highly sensitive to the order in which tokens are unmasked during the denoising process.
• →A learned scheduler using KL-regularized MDP framework replaces traditional heuristic-based unmasking schedules.
• →The optimized policy generates samples that more closely match data distributions than existing heuristic methods.
• →Empirical results show consistent outperformance across four benchmarks, with particularly strong gains on the SUDOKU dataset.
• →The research provides theoretical guarantees for policy improvement and convergence under standard assumptions.