YKSBench

Large Multimodal Models (LMMs) have achieved rapid progress in bridging language and vision, yet their performance on visually grounded, exam-style reasoning tasks remains far from human-level. YKSBench addresses this gap by providing a multimodal benchmark constructed from 2,047 Turkish university entrance exam (YKS) preparation questions spanning mathematics, geometry, physics, chemistry, biology, and geography.

Unlike prior benchmarks that focus narrowly on mathematics or synthetic diagram tasks, YKSBench captures diverse question formats where visual grounding is indispensable. Despite their apparent simplicity to students familiar with YKS, state-of-the-art LMMs exhibit systematic and surprising failures. In our evaluations, the best-performing proprietary model, GPT-5, achieved only 39.34% accuracy, while the strongest open-source model, Gemma 3 27B, reached 26.82%.

Our analysis includes qualitative case studies and scientific visualizations illustrating recurring failure modes, including misinterpreting diagrams, misreading schematics, and overlooking spatial relationships. As a complementary resource to MathVista, MathVision, and MMStar, YKSBench expands the evaluation landscape and highlights the urgent need for LMMs with more robust visual reasoning capabilities.

Link to paper

---

Dataset Details

Dataset Description

Developed by Middle East Technical University (METU) Computer Engineering, YKSBench is a multimodal benchmark built from 2,047 YKS university entrance exam preparation questions. The dataset spans six major STEM domains and contains a rich variety of diagram-based and visually grounded question types.

Each problem includes:

• A high-resolution diagram or visual stimulus
• Associated textual question content
• Gold answer labels

YKSBench is designed for academic benchmarking with a focus on evaluating visual perception, diagram understanding, and multimodal reasoning. For academic research use only.

• Curated by: Egemen Sert, Şeyda Ertekin
• Language(s): Turkish, English
• License: MIT

---

Uses

YKSBench is intended exclusively for research and benchmarking in multimodal reasoning. Recommended uses include:

• Evaluating diagram understanding in LMMs
• Stress-testing visual reasoning under realistic exam constraints
• Measuring cross-domain reasoning robustness
• Studying failure modes in perception-grounded reasoning tasks

Commercial use is not allowed.

---

Citation

If you use YKSBench, please cite:

BibTeX:

@misc{yksbench2025,
  title        = {YKSBench: Stress-Testing Multimodal Models with Exam-Style Questions},
  author       = {Sert, Egemen and Ertekin, Şeyda},
  year         = {2025},
  howpublished = {\url{https://openreview.net/forum?id=qgAjoo3cJE}},
  note         = {Accessed: 2025-12-01}
}

---

Dataset Card Contact

For submissions, issues, and benchmark-related questions, please contact: [email protected]