MathScrub: Single-Pipeline Strikeout Removal and LaTeX Generation for Handwritten Calculus Expressions

With the rapid development of large language models (LLMs), image-to-LaTeX conversion has shown great potential for applications such as automated grading and mathematical expression recognition in educational and assessment settings. However, real-world handwritten STEM manuscripts often con tain strike-outs, overwrites, and other markings that disrupt handwritten content structure, reducing recognition accuracy and compilability. We present MathScrub, a single-pipeline multimodal framework that cleans and transcribes handwritten calculus expressions by grouping connected components with geometry-constrained clustering, detecting deletions via an early fusion Vision Transformer, applying geometry-aware inpainting, and generating compilable LaTeX with a parameter-efficiently adapted open-source multimodal VLM. To support this task, we introduce Math-Strike, a dataset of handwritten calculus answers with real and synthetic strike-outs, component-level annotations, and aligned LaTeX. Experiments show MathScrub outperforms open-source baselines on component detection and image-to LaTeX tasks, while remaining robust in real strike-out scenarios, providing a reliable pre-processing stage for LLM-based grading and formula recognition.