Megadose AI progress, ranked and analyzed.

Simplex Relaxation for Discrete Diffusion

· HF Daily Papers ·
Simplax adds a simplex-valued auxiliary variable to uniform discrete diffusion while keeping the original categorical corruption process intact.

The paper frames this as an exact Dirichlet-categorical augmentation, not a replacement for the diffusion kernel. It gives a tractable Rao-Blackwellized reverse-bridge objective and a stochastic reverse sampler, with the denoiser still seeing the corrupted categorical state. In reported tests, Simplax improves the perplexity-entropy tradeoff on unconditional OpenWebText generation. On Sudoku, training only on 30-clue puzzles produced the top accuracy among compared methods across evaluated clue densities, including 17-clue puzzles, and the highest unconditional validity. HF Daily Papers' note

score 4

Categories: Research