Calibrated typed-decision models, sometimes called System One models, return a label and a probability in a single forward pass, making them far cheaper than large language models. But their evaluation has been almost entirely English, leaving open the question of whether their calibration and accuracy transfer to other languages.
This paper takes up that question for Indonesian regional languages. The authors examine how label budgets and translation affect performance, asking whether the efficiency of System One models survives outside the English-centric benchmark setting.
If the models transfer well, they could offer a low-cost path to classification in under-resourced languages. If not, the results would point to the need for language-specific calibration data rather than assuming English results generalize.