The integration and application of Multimodal Large Language Models (MLLMs) in the education sector offers significant potential for automating tasks such as exam grading and personalized learning. However, current MLLMs are predominantly trained on English text and simplified document formats, which limits their performance on complex, non-English documents like exam papers. Key challenges in understanding exam papers include accurate text extraction, maintaining the correct reading order, and ensuring logical coherence across multiple pages. Addressing these issues is critical for the practical deployment of MLLMs in educational applications. To advance research in this area, we organized the ICDAR 2025 Competition on Understanding Chinese College Entrance Exam Papers. This competition introduced the CEP-7K dataset, consisting of 7,000 question-answer pairs and their corresponding exam paper images, to assess MLLMs’ abilities in tackling these challenges. Held from December 10, 2024, to April 18, 2025, the competition received six valid submissions. Compared to the benchmark model Qwen2.5-VL-32B-Instruct, the participating teams have achieved up to a 42% improvement in accuracy. This report provides an overview of the dataset, task design, and evaluation metrics. It also presents the evaluation results and highlights key insights and advancements gained from the competition.