Interactive Demo · Sudoku Hard

Visualizing Diffusion Language Model Reasoning

Comparing our modified vs default diffusion language model training objective. By also noising the prompt, we expand the range of prediction tasks, leading to greater generation diversity and reasoning performance. Full details: link.

step 0 / 0

Sampled from a random set of 128 puzzles. Both models use the same seeded noise for fair comparison. Puzzle turns green when correct, red highlighted cells are incorrect cells.