Research
Research
Papers, evaluations and methods from NoemaAI and our collaborators. Our published work so far covers post-training small language models, evaluating them against their base models, running them efficiently on personal devices, and aggregating model probabilities.
Areas
Each area links to the published work behind it.
Post-training small models
Noema 2B, Noema 1.5 2B and Noema 1.5 4B are post-trained from Qwen3.5 bases and released as open weights, each with a full evaluation on its model page.
Evaluation methodology
The Noema 1.5 4B evaluation pairs every comparison with its base model on identical items, adds a held-out stress test and 95% intervals, and reports losses alongside gains.
Efficient on-device inference
Overfit runs mixture-of-experts models larger than device memory by streaming experts from storage on demand. It is experimental.
Probability aggregation
When summing fine-grained model probabilities into broader categories improves predictions, and when it only guarantees consistency.
Publications
Newest first. Each entry has its own page with the abstract, a citation and links to the paper.
- Preprint
Reusing Model Probabilities for New Queries
Can one set of model probabilities answer many different questions? This paper studies how fine-grained predictions can be combined into broader categories, keeping answers consistent across label groupings without querying a model again for each grouping.
- Probability aggregation
- Generalization
- Semantic parsing
PDF of Reusing Model Probabilities for New QueriesDOI 10.5281/zenodo.23584994Zenodo
Follow the research
New papers and research notes are announced in the newsroom and in the research feed.