Skip to content
PRAXNIADIRGACAKSU NUSANTARA
Try Kalnavi

Research library

The papers and studies Kalnavi is built on. We read them, test them, and apply them with care.

  1. 1984

    How we use itThe basis of our mission: one-to-one tutoring is the most effective way of learning ever measured.

  2. 2006
    Test-Enhanced Learning: Taking Memory Tests Improves Long-Term Retention

    Roediger, H. L., & Karpicke, J. D. — Psychological Science, 17(3), 249–255

    How we use itKalnavi uses quizzes and retrieval practice, not just rereading.

  3. 2006
    Distributed Practice in Verbal Recall Tasks: A Review and Quantitative Synthesis

    Cepeda, N. J., Pashler, H., Vul, E., Wixted, J. T., & Rohrer, D. — Psychological Bulletin, 132(3), 354–380

    How we use itPractice is scheduled again and again with gaps, so knowledge lasts.

  4. 2023
    ChatGPT for good? On opportunities and challenges of large language models for education

    Kasneci, E., et al. — Learning and Individual Differences, 103, 102274

    How we use itA map of opportunities and risks of language models in education that we use as a checklist.

  5. 2023
    Guidance for generative AI in education and research

    UNESCO (Miao, F., & Holmes, W.) — UNESCO

    How we use itGuidance on age, privacy, and human oversight for generative AI in schools.

  6. 2017
    Attention Is All You Need

    Vaswani, A., et al. — NeurIPS 2017

    How we use itThe Transformer architecture, the foundation of every language model we use.

  7. 2020
    Language Models are Few-Shot Learners

    Brown, T. B., et al. — NeurIPS 2020

    How we use itShowed that large language models can learn new tasks from a few examples.

  8. 2022

    How we use itPrompting models to reason step by step, like a teacher explaining at the board.

  9. 2021

    How we use itAn efficient way to adapt large models to Indonesian language and curricula.

  10. 2022

    How we use itTraining models to follow human instructions and preferences.

  11. 2022

    How we use itAligning models with a written set of principles that can be inspected.

  12. 2023

    How we use itThe preference alignment method we use for the Kalnavi model.

  13. 2024

    How we use itIntroduced GRPO, which we use to train mathematical reasoning.

  14. 2020

    How we use itAnswering from retrieved sources, so answers can be checked.

  15. 2023

    How we use itThe think-then-act pattern behind Kalnavi agents.

  16. 2023

    How we use itModels that learn to use tools such as calculators and search.

Our research agenda

  1. 1

    Language models for Indonesia

    Kalnavi-8B, a model we further train with SFT, DPO, and GRPO for Indonesian language and curricula. In development.

  2. 2

    Measuring real learning

    Evaluations that test whether students truly understand more, not just whether they got an answer.

  3. 3

    Child safety

    Age limits, content filtering, and human oversight that can be audited.