Khanh-Tung Tran, Barry O'Sullivan, Hoang D. Nguyen

Accepted to AAAI-26

This repository contains code and resources for the paper Reasoning Alignment for an Extremely Low-resource and Endangered Language: Separating Reasoning and Language Understanding.

• ./data: evaluation data, including our contributed dataset LC2024.
• ./src/train: training code, including for the baseline Native-CoT Training and English-Pivoted CoT Training.
• ./src/evaluation: evaluation code, based on the SkyThought repo.
• ./appendix.pdf: technical appendix.

1. Install dependencies:

   cd ./src/train
   pip install -r requirements.txt

2. Run the main script:

   bash scripts/lang_adapt/run_sft.sh # change the paths accordingly

1. Install dependencies:

   cd ./src/evaluation
   pip install -r requirements.txt

2. Run evaluation:

   cd ./src/evaluation/skythought/skythought_evals
   python eval.py --model ${YOUR_MODEL_HERE}$ --evals=aime,irish_aime,LC2024 --tp=1 --output_file=results.txt --temperatures 0.6 --n 64

For more information, refer to the original document from the SkyThought repo.

@article{Tran2026,
  title = {Reasoning Transfer for an Extremely Low-Resource and Endangered Language: Bridging Languages Through Sample-Efficient Language Understanding},
  volume = {40},
  ISSN = {2159-5399},
  url = {http://dx.doi.org/10.1609/aaai.v40i39.40613},
  DOI = {10.1609/aaai.v40i39.40613},
  number = {39},
  journal = {Proceedings of the AAAI Conference on Artificial Intelligence},
  publisher = {Association for the Advancement of Artificial Intelligence (AAAI)},
  author = {Tran,  Khanh-Tung and O’Sullivan,  Barry and Nguyen,  Hoang D.},
  year = {2026},
  month = mar,
  pages = {33277–33286}
}