Sudoku Diffusion

A masked-diffusion language model (iLLaDA-8B) fine-tuned with LoRA to solve 9×9 Sudoku, compared against autoregressive LLMs. Diffusion fills the whole grid in parallel and unmasks the cells it is most confident about first, so it can use constraints from anywhere on the board instead of committing left to right.
Adapter: Techno03/illada-8b-sudoku-lora · Code: github.com/arush-garg/SudokuDiffusion

Benchmark

Difficulty	iLLaDA zero-shot	Gemma-4 zero-shot	iLLaDA fine-tuned	Gemma-4 fine-tuned	Llama 3.3 70B
easy (n=13)	~35.8%	72.5%	92.7%	95.7%	35.2%
medium (n=70)	~8.2%	21.5%	69.6%	38.9%	14.4%
hard (n=17)	~0.0%	3.6%	27.4%	8.9%	7.5%

Metric: % of originally-blank cells predicted correctly, on the same 100 held-out validation puzzles. iLLaDA zero-shot numbers are estimates from a small sample.

Compare the models puzzle by puzzle

These are the real outputs recorded during evaluation for the 100 held-out puzzles, scored cell by cell. Live inference isn't hosted here because the 8B model needs a GPU.

Difficulty

Puzzle

givencorrectwrongno parseable answer