Noma: An Open System One Decision Model from a Sliced Decoder and Trained Heads
Atul Saxena and Divyanshi Sharma. Blackdrome AI Labs.
What the paper shows.
Noma is an open-weight System One decision model: it reads a state and typed questions and returns a probability for every option and a separate abstain signal in one forward pass, without generating text. The paper documents the architecture and tests it with 20 controlled training runs.
The first 18 of the backbone's 32 layers score the same as all of them on the sealed set: 79.4% against 79.4%, over three seeds each.
The separate abstain output is informative only when it is trained with its own supervision.
One small pointwise head, about an eighth the size of the released four-head design, scores within two points of it.
On held-out questions that need chained computation, every configuration stays near half. The paper reports this alongside the results.
BibTeX.
@misc{saxena2026noma,
title = {Noma: An Open System One Decision Model from a Sliced Decoder and Trained Heads},
author = {Saxena, Atul and Sharma, Divyanshi},
year = {2026},
publisher = {Zenodo},
doi = {10.5281/zenodo.23186353},
url = {https://doi.org/10.5281/zenodo.23186353}
}DOI: 10.5281/zenodo.23186353. For the model itself, see the Noma overview.