Researchers have introduced OmniHandwritingOCR, a benchmark for evaluating the ability of large language models to recognize handwritten text. The benchmark covers six subtasks and twelve subsets, including recognizing multilingual handwriting, writer errors, and complex mathematical expressions. It consists of over 77,000 labeled images from public datasets and student writings. Current systems struggle with faithful transcription, particularly on complex formulas, and often hallucinate corrections. OmniHandwritingOCR provides a diagnostic tool for identifying the limitations of multimodal models in handwritten OCR scenarios. --- Why it matters: This benchmark matters to researchers because it highlights the challenges of recognizing handwritten text, which is still a difficult task for AI systems. Understanding these limitations is crucial for developing more accurate and robust OCR systems that can handle real-world documents. Source: https://arxiv.org/abs/2608.18586