A 1.28 M-parameter Diffusion Transformer trained on only 1,000 Sudoku-Extreme
puzzles — the same small-data regime HRM and TRM report on. It solves Sudoku as a
masked discrete diffusion denoiser: fill the cells you are most confident about,
re-read the board, repeat.
watch it solve
This model solving a 51-blank board over 24 reveal steps. Given clues stay grey; filled
cells are tinted by the model's step-0 confidence (blue = unsure -> green = sure).
Architecture: DiT with adaLN-Zero conditioning — hidden=128, heads=4,
blocks=4; per-cell token + 2-D positional + 3x3-box embeddings, plus a timestep.
Sibling model:tchauffi/sudoku-dit,
same architecture trained on a 3.8 M-board stream of ordinary Sudoku.
A board is 81 tokens, row-major: 0 = empty / [MASK], 1..9 = digits. The solver
clamps the given clues and only fills the blanks.
Usage
python
1import torch
2from nonet.hub import load_solver # pip install git+https://github.com/tchauffi/nonet34solver = load_solver("tchauffi/sudoku-dit-extreme-1k")# the config records the cosine-high schedule5puzzle ="402609001000000000700002084000080500030005007004060930008021406005000000147030000"6x = torch.tensor([[int(c)for c in puzzle]])# (1, 81), 53 blanks7solution = solver.solve(x, conf_threshold=0.999)# adaptive reveal89"".join(map(str, solution[0].tolist()))10# 482659371613874295759312684971483562836295147524167938398721456265948713147536829
This is also the default model for load_solver() with no arguments.
Performance
Held-out Sudoku-Extreme test puzzles, adaptive decoder (tau = 0.999). Because every
solution is exactly checkable, a wrong answer can simply be retried: attempt 1 is the
greedy decode, later attempts resample at temperature 1 (randomness enters only at the
forced below-tau reveals). k is the attempt budget.
attempts k
overall
rating <= 2
rating 3-9
rating 10+
1 (greedy)
22.7 %
75.2 %
11.1 %
3.9 %
8
33.1 %
86.1 %
21.5 %
14.1 %
32
49.7 %
95.6 %
42.1 %
32.8 %
64
60.4 %
97.6 %
53.8 %
46.8 %
For reference, HRM reports 55.0 % on this benchmark at 27 M parameters (~21x
larger). The comparison is not compute-matched: our 60.4 % spends 64 decode attempts per
puzzle, and the curve is still unsaturated at k = 64. The honest reading is that a tiny
diffusion denoiser plus a verifier covers this regime, not that it is cheaper per puzzle.
Training
Data: 1,000 puzzles from
sapientinc/sudoku-extreme,
plus the Sudoku symmetry group sampled per example per batch (digit relabeling, transpose,
band/row and stack/column permutations) — validity-preserving, the same family HRM/TRM use.
Objective: masked cross-entropy over masked cells, conditional (given clues are never
masked), cosine-high masking schedule (mass where deduction is hard).
30 k steps, AdamW, batch 256, lr 3e-4.
Augmentation is load-bearing. The identical run on the raw 1,000 boards without symmetry
augmentation memorizes them and collapses to 0.1 % test-valid. With augmentation the
effective board count is large enough to generalize.
Limitations
Not a guaranteed solver: the hardest rating bands remain below 50 % even at k = 64.
The headline number depends on a verifier and a 64x decode budget; single-shot accuracy is
22.7 %.
