Part 4 of 8

The Maths II: Masked Discrete Diffusion

Everything in the previous section assumed you can add and scale — operations tokens don’t support. This section rebuilds the pipeline for discrete data, following the notation of the MDLM paper (Kuleshov’s lab), which is the cleanest formulation and the one the later papers build on. Time now runs continuously from t = 0 (clean) to t = 1 (fully corrupted).

1. From Gaussians to categoricals
2. The reverse posterior
3. The objective: BERT, integrated
4. Where the speed–quality trade-off lives, formally
5. Quiz: 5 questions