Raghuram Shankar
← Diffusion Language Models

Part 5 of 8

The Maths III: Inside the SOTA Methods

The Papers & Sources part tells the story of the modern diffusion-LM methods; this section shows their actual mathematics. Each of the four ideas below modifies exactly one piece of the Maths II machinery — the factorization, the sampler’s kernel, the per-step distribution, or the training signal.

  1. Block Diffusion: the block-factorized bound
  2. ReMDM: the remasking kernel, with its feasibility bound
  3. Discrete guidance: sharpening by ratio
  4. diffu-GRPO: policy gradients without left-to-right log-probs
  5. Quiz: 5 questions