Truth or Trick? A Vietnamese Dataset of Adversarial Logic in Geometry

Multimodal mathematical reasoning demands that large language models jointly parse diagrams and text, yet their robustness to adversarial logic—particularly in low-resource languages—remains unclear. To probe this gap we introduce ViGeoTrap, the first Vietnamese benchmark for adversarial reasoning over geometric diagrams. Each of its 635 items pairs a diagram with a Vietnamese statement that is either correct or minimally contradicted, requiring a TRUE/FALSE decision from the image alone. Three fine-tuned models (ViLT, CLIP, BLIP) and four zero-shot LLaMA-3.2/4 variants are evaluated: ViLT leads among fine-tuned systems, BLIP struggles on truthful cases, and LLaMA-4 Maverick and Scout show the strongest zero-shot robustness. The results underline how architecture (e.g., early fusion) and training objectives (e.g., instruction tuning) govern adversarial performance in low-resource settings. ViGeoTrap is publicly available at https://huggingface.co/datasets/longndh/ViGeoTrap.