Recursive AI Grading Assistant

Education systems in many countries still rely heavily on paper-based homework and tests, especially in elementary and middle schools. Grading these materials manually is time-consuming and adds to teachers' already heavy workloads. Digitalizing and automating grading is key to improving efficiency and enhancing the quality of education.

Recursive developed an advanced AI Grading Assistant that automatically evaluates elementary school students’ handwritten math answers. The assistant supports core mathematical functions, including addition, subtraction, multiplication, and division with both whole numbers and decimals, as well as linear (horizontal) and column (vertical) calculations.

Using this tool, students can upload photos of handwritten homework and tests through a user-friendly web app accessible on PCs, tablets, and smartphones. The assistant intelligently interprets the handwritten input, converting it into a machine-readable format for analysis, computes correct solutions, and provides instant score results.

Highlights

• High Accuracy Results: Recursive’s engineering team collected and labeled the vast amount of handwritten math answers to train the model. This extensive dataset significantly enhanced its performance, resulting in an average accuracy of 85%.
• Support for Complex Math Problems: In addition to whole numbers and linear (horizontal) calculations, the assistant supports decimals and column (vertical) calculations, which have traditionally been challenging for AI due to technical limitations. Recursive also explores expansion to middle school-level problems, including equations and geometry.
• Automated Grading in Just 10 Seconds: The assistant automatically evaluates answers in under 10 seconds, saving teachers time and allowing them to focus on lesson planning and personalized teaching activities.
• Compliance with Educational Standards: The assistant can be fine-tuned to follow specific scoring guidelines, such as those set by government education ministries. It ensures that answers meet required formats, aligning with standardized grading criteria.
• Secure & Privacy-Focused Infrastructure: The assistant operates fully on Recursive’s secure GPU infrastructure, ensuring that data remains secure and fully controlled by the client.
• Web-Based Application for Easy Access: Teachers can access evaluation results via a web-based platform compatible with PCs, tablets, and smartphones. This allows for efficient student performance tracking and streamlined grading management.

Implementation

• Data sources:

The main dataset includes a vast amount of handwritten math answers, covering a variety of handwriting styles, problem types, and answer formats. This diversity helps ensure the model’s robustness and generalizability.

• Data Labeling:

The Recursive team refined the labeling process through iterative improvements and quality control measures. This significantly enhanced the model’s ability to correctly identify and annotate key areas within the handwritten answers.

A hybrid approach, combining automated labeling with human review and corrections, ensured high-quality labeled data by eliminating errors and inconsistencies from the automated process.

• AI model:

Recursive leverages a custom-built Large Language Model (LLM) with math-optimized OCR capabilities to process and understand handwritten math problems. This model is designed to both identify the relevant components within an image of a handwritten problem and accurately transcribe the handwritten content into a machine-readable format.

• Result Evaluation:

The LLM model outputs a string representing the transcribed answer. This string is then systematically evaluated using predefined rules and algorithms to determine the correctness of the answer. This evaluation process takes into account various factors, such as numerical accuracy, adherence to mathematical principles, and consistency with expected answer formats.

• Business Outputs:

The assistant is delivered via a user-friendly web-based platform, providing teachers with easy access to student performance data and insights. By automating the grading process, Recursive AI Grading Assistant saves teachers a significant amount of time, allowing them to focus on more strategic activities such as lesson planning, personalized teaching, and student engagement.