VN-HSGExam: A Vietnamese High School Graduation Examination Benchmark for LLM Evaluation

We introduce VN-HSGExam, a benchmark of 818 text-evaluable multiple-choiceitems from the official Vietnamese High School Graduation Examination(2019-2024) across four subjects (History, Geography, Civics, Biology). Weevaluate four open-weight LLMs (Vistral 7B, Llama 3.1 8B, GPT-OSS 120B,Qwen3 235B) under four prompting strategies, yielding 16 conditions over8,256 inference calls on a balanced 516-item Tier 1 subset. Our findings:(1) scale dominates specialization (Qwen3 235B exceeds Vistral 7B by 15.9percentage points); (2) Civics is uniformly the hardest subject across allmodels, with peak accuracy 38-44%; (3) prompting strategy effects arestatistically significant only for sub-10B models (McNemar p = 0.0002 forVistral 7B).