🧮📐 MAGR: Towards Automatic Rubric-Based Grading for Mathematics ✏️✨

🎉 We present MAGR (Math Auto Grading with Rubrics), a benchmark introduced in the paper "MAGR: Towards Automatic Rubric-Based Grading for Mathematics". MAGR is designed to evaluate fine-grained step-level grading and visual grounding capabilities of Vision-Language Models (VLMs) on real-world handwritten math problems! 📝🧠

Developed upon the foundational FERMAT benchmark, our proposed MAGR advances from coarse solution-level evaluation to granular step-level rubric grading paired with dense visual grounding.

---

🎯 What’s Inside MAGR?

1,968 handwritten math solutions derived from 529 unique problems, carefully annotated across core mathematical domains: 🔢 Algebra | 📐 Geometry | 📊 Matrices & Determinants | 📐 Trigonometry | 📈 Calculus | 🎲 Probability & Statistics

Each annotated instance features:

1. 📝 Granular Step-Level Rubrics: Student steps are mapped against teacher rubrics with boolean correctness (is_correct) and diagnostic explanations (reasoning).
2. 🎯 Dense Visual Grounding: Every intermediate step is grounded with bounding boxes (bbox) and polygon masks (segmentations).
3. 🤖 Dual Model Perspectives: Parallel evaluation splits generated using state-of-the-art Gemini-3.1-Pro and GPT-5.6-Terra, perfect for studying model consistency and LLM-as-a-judge behaviors.

🖼️ Visual Grounding Example

Error Explanation: The student misapplied the reduction formula by confusing sin with cos . The red bounding box precisely pinpoints the exact location of this error on the handwritten response.

---

📦 Dataset Features

Each instance in MAGR consists of:

Main Content

• problem: Original math problem prompt (LaTeX)
• student_steps: Step-by-step breakdown of the student's solution (S1, S2, ...)
• teacher_rubrics: Reference teacher solution steps / rubrics (T1, T2, ...)
• step_evaluations: Evaluation per step (teacher_step_id, is_correct, reasoning)
• annotations: Dense visual grounding (teacher_step_id, segmentations, bbox)

Metadata

• filename: Relative path to the handwritten image (images/<new_custom_id>.jpg)
• img_width: Width of the image (in pixels)
• img_height: Height of the image (in pixels)
• problem_id: Base problem ID
• new_custom_id: Unique sample identifier

---

🚀 Why Use MAGR?

MAGR is designed for evaluating fine-grained reasoning and auto-grading capabilities of Vision-Language Models (VLMs) on handwritten math:

• ✅ Step-Level Rubric Alignment: Diagnoses reasoning flaws even when final answers happen to be correct.
• ✅ Dense Visual Grounding: Pinpoints exact coordinates of handwritten mathematical steps.
• ✅ Dual-Model Perspectives: Benchmarked across two leading paradigms (Gemini-3.1-Pro & GPT-5.6-Terra).

---

📖 Citation

If MAGR helps your research, kindly include the following citations:

@article{luc2026magr,
  title   = {MAGR: Towards Automatic Rubric-Based Grading for Mathematics},
  author  = {Luc Do Quang and Collaborators},
  year    = {2026},
  journal = {arXiv preprint}
}

@article{nath2025vision1language,
  title   = {Can Vision-Language Models Evaluate Handwritten Math?},
  author  = {Oikantik Nath and Hanani Bathina and Mohammed Safi Ur Rahman Khan and Mitesh M. Khapra},
  year    = {2025},
  journal = {arXiv preprint arXiv: 2501.07244}
}