LaTA: A Drop-in, FERPA-Compliant Local-LLM Autograder for Upper-Division STEM Coursework
Innovations in educational technology continue to redefine the landscape of learning and assessment in academia. Among the latest advancements is LaTA, an open-source autograder designed to streamline the grading process for upper-division STEM courses while ensuring compliance with the Family Educational Rights and Privacy Act (FERPA). This new tool addresses critical concerns associated with traditional grading methods, particularly the risks associated with sending student data to third-party applications.
LaTA, which stands for LaTeX Teaching Assistant, operates entirely on local, on-premises hardware, making it a secure and efficient option for educational institutions. The system is built on a four-stage pipeline: ingest, segment, grade, and report. This sequence allows LaTA to analyze student submissions effectively, utilizing a locally hosted open-weight chain-of-thought large language model (LLM) grader, specifically gpt-oss:120b. One of the standout features of LaTA is its ability to work seamlessly with a LaTeX-native workflow, which is already commonplace in many engineering and physics courses.
Key Features of LaTA
- FERPA Compliance: By processing student data on local servers, LaTA eliminates the risks associated with third-party data handling, ensuring compliance with federal privacy regulations.
- Cost-Effective: The deployment of LaTA in Winter 2026 at Oregon State University demonstrated a marginal cost of $0 per assignment, making it highly sustainable for institutions.
- Time Efficiency: Grading time was significantly reduced to just 1-3 minutes per submission, freeing up valuable time for instructors and teaching assistants.
- Improved Learning Outcomes: In a comparative study, students graded by LaTA outperformed their traditionally graded peers by approximately 11% on midterm exams and 8% on final exams.
- Enhanced Student Confidence: A survey of 159 students indicated significant improvements in self-assessed confidence across all stated learning objectives, with an average increase of 1.49 Likert points.
The practical application of LaTA was tested in the Mechanical Engineering Methods course (ME 373), which involved approximately 200 students. The instructor noted an impressively low grading error rate of 0.02% to 0.04% per rubric line item throughout the term. This accuracy, coupled with the rapid grading turnaround, enabled instructors to offer more comprehensive feedback and increased office hours, enhancing the overall academic experience for students.
The results from the LaTA implementation not only highlight the potential for improved educational assessments but also underscore the importance of maintaining data privacy in an increasingly digital learning environment. As educational institutions continue to seek effective solutions for managing grading workloads, LaTA presents a promising alternative that balances efficiency, compliance, and academic integrity.
LaTA is now available to the educational community under the AGPLv3 license, allowing institutions to implement and customize the tool according to their specific needs. As more educators adopt local LLM-based grading systems, the potential for enhancing student learning experiences while safeguarding sensitive information will undoubtedly grow.
Related AI Insights
- HiMAC: Hierarchical Learning for Long-Horizon LLM Agents
- Closed-Loop Vision-Language Planning for Multi-Agent AI
- Improving AI Safety with Annotator Policy Models
- Flow Sampling: Efficient Sampling from Unnormalized Densities
- Deep Learning Advances in Photoplethysmography Analysis
- BALAR: Bayesian Loop Enhances AI Active Reasoning
- Sycophancy in LLMs: Balancing Helpfulness & Integrity
- Inconsistent Databases & Argumentation Frameworks with Collective Attacks
- Ensuring Safety Before Deploying Open-Ended AI Systems
- AI and Human Collaboration for Smarter Inventory Control
