<ul><li>Recent research introduces EB-Sampler, a method for efficient sampling from masked diffusion models (MDMs) used for language modeling.</li><li>EB-Sampler utilizes an Entropy Bounded unmasking procedure to predict multiple unknown tokens with a single function evaluation with predefined error tolerance.</li><li>The new sampling method accelerates sampling from MDMs by 2-3x on standard coding and math reasoning benchmarks, with no loss in performance.</li><li>EB-Sampler also shows effectiveness in smaller reasoning tasks like maze navigation and Sudoku, tasks where autoregressive models often struggle.</li></ul>

Accelerated Sampling from Masked Diffusion Models via Entropy Bounded Unmasking

Discover more