FERMAT: Can Vision-Language Models Evaluate Handwritten Math?

📜 Paper | 🤗 HF Dataset

We present FERMAT, a benchmark designed to assess VLMs’ ability to detect, localize and correct errors in handwritten mathematical content. Please refer to our paper for more details.

VLMs

• Gemini Family
• GPT Family
• Pixtral Family
• Llama Family
• Phi3.5V

Loading Data

Steps to download data and store the images in benchmark_images, and csv in benchmark_csv. Steps to dowload data for the oikantik format

Setup

To run evaluation of VLMs against the FEMRAT dataset, you need to install the required packages by running the following command:

pip install -r requirements.txt

We self-hosted Pixtral-12B-2409 (https://huggingface.co/mistralai/Pixtral-12B-2409), Pixtral-Large-Instruct-2411, LLaMa-3.2-11B-Vision-Instruct, LLaMa-3.2-90B-Vision-Instruct, Phi-3.5-Vision-Instruct using VLLM (/~https://github.com/vllm-project/vllm)

We used hosted services for GPT-Family, Gemini-Family

For self-hosted models,

1. Set up environment variables:

   export OPENAI_API_BASE=[ADD_THE_ENDPOINT_URL_OF_HOSTED_MODEL]

   Example: "http://localhost:8004/v1"

2. Start Evaluations:

   python main.py --model [MODEL_NAME] --dir_name [DATA_DIR]

   • MODEL_NAME: Name of the model to be evaluated. Choices: ['pixtral', 'pixtral_large', 'phi', 'llama_large', 'llama']
   • DATA_DIR: Path to the directory where the Benchmark Images are stored

3. Fill-in CSV

   Once the evaluation is done, the results will be stored in a JSON File with the format state_<MODEL_NAME>.json. You can convert this JSON file to a CSV file using the following command:

   python fill_in_csv.py --model [MODEL_NAME] --csv-file [CSV_FILE] --json-file [JSON_FILE]

   • MODEL_NAME: Name of the model to be evaluated. Choices: ['pixtral', 'pixtral_large', 'phi', 'llama_large', 'llama']
   • CSV_FILE: Path to the CSV file where the results need to be filled in.
   • JSON_FILE: Path to the JSON file where the results are stored.

Evaluation

Error Detection

Error Localization

Error Correction

Citation

If you used this repository or our models, please cite our work:

@misc{nath2025visionlanguagemodelsevaluatehandwritten,
      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},
      eprint={2501.07244},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2501.07244},
}