Researchers have released OpenTWBench, an open evaluation suite designed to test large language models' knowledge of Taiwan-specific domains. The suite consists of two benchmarks, the largest of which, tw-exam-bench, contains 255,854 multiple-choice questions spanning 148 academic subjects.
The suite aims to fill a gap in evaluation resources for Taiwan-domain knowledge. In addition to presenting the benchmarks, the paper investigates how the construction of the measurement apparatus influences the scores that models receive, suggesting that benchmark design choices are not neutral.
Because this is a single source, there are no independent findings to compare. The abstract does not name the second benchmark or describe the exact origin of the exam questions beyond the 148-subject scope.