Binary Discrete Diffusion: models
The models of Binary Discrete Diffusion: Candidate-Set Diffusion for Logic and Planning
Tasks: diffusion transformers for each task and noise kernel (masked,
uniform, binary_symmetric, binary_absorbing), five training seeds on the logic tasks and one on text. Each
run directory of binary-discrete-diffusion was trained
with configs/paper/<task>/<kernel>.yaml and holds the composed config in .hydra/config.yaml and the EMA
parameters the paper evaluates in checkpoints_ema/<step>/. The logic runs also hold their final training state
(parameters, optimizer and time-sampler state) in checkpoints/<step>/, from which train.py resumes.
| Task | Directory | EMA checkpoint | Final training state | Data |
|---|---|---|---|---|
| Sudoku | sudoku/<kernel>/seed<0-4> |
checkpoints_ema/1000000 |
checkpoints/1000000 |
scripts/prep_sudoku.py |
| Boxoban | boxoban/<kernel>/seed<0-4> |
checkpoints_ema/2500000; binary absorbing seeds 1 and 3 at 2000000 and 1000000 |
checkpoints/2500000 |
scripts/prep_boxoban.py |
| Graph coloring | graphcolor/<kernel>/seed<0-4> |
checkpoints_ema/300000 |
checkpoints/300000 |
generated on the fly |
| Text8 | text8/<kernel> |
checkpoints_ema/1000000 |
characters, written by scripts/prep_text8.py |
|
| TinyStories | tinystories/<kernel> |
checkpoints_ema/500000 |
tokenizer rodcasnog/tinystories-bpe-1k | |
| OpenWebText | owt/<kernel> |
checkpoints_ema/250000 |
tokenizer dvruette/openwebtext-bpe-1k |
The Boxoban EMA checkpoints were selected on validation NELBO, as described in the paper's appendix. Since
train.py resumes only from a step that holds both checkpoints, Boxoban binary absorbing seeds 1 and 3 do not resume
as downloaded.
Usage
From a clone of the code repository, after uv sync:
uv run hf download rodcasnog/binary-discrete-diffusion --include "sudoku/binary_symmetric/seed0/*" --local-dir outputs/hf
uv run python -m binary_diffusion.eval.solve checkpoint_dir=outputs/hf/sudoku/binary_symmetric/seed0/checkpoints_ema/1000000 \
sampler.nfe=256 out=outputs/sudoku-bsym/solve
# Train the same run 200k steps longer.
uv run python -m binary_diffusion.train --config-name paper/sudoku/binary_symmetric training.seed=0 \
training.num_steps=1200000 hydra.run.dir=outputs/hf/sudoku/binary_symmetric/seed0
uv run hf download rodcasnog/binary-discrete-diffusion --include "text8/binary_symmetric/*" --local-dir outputs/hf
uv run python -m binary_diffusion.sample checkpoint_dir=outputs/hf/text8/binary_symmetric/checkpoints_ema/1000000 \
sampler.nfe=256 sampler.spacing=log sampler.eta=1 sampler.lam=1 out=outputs/text8-bsym/samples
Sudoku and Boxoban models read their task data in data/sudoku and data/boxoban. Text8 models read the tokenizer
in data/text8, TinyStories samples take their lengths from data/tinystories, and NELBO and MAUVE read the corpus;
the code's README lists the data and evaluation commands.
License
Apache-2.0.
Citation
@misc{casadonoguerales2026binarydiffusion,
title = {Binary Discrete Diffusion: Candidate-Set Diffusion for
Logic and Planning Tasks},
author = {Casado Noguerales, Rodrigo and von R{\"u}tte, Dimitri
and Sch{\"o}lkopf, Bernhard and Hofmann, Thomas
and Raoufi, Aran},
year = {2026},
eprint = {2610.28543},
archivePrefix = {arXiv},
primaryClass = {cs.LG},
url = {https://arxiv.org/abs/2610.28543},
}