LaTA: Open-Source Local LLM Autograder for STEM Courses
LaTA (LaTeX Teaching Assistant) serves as an open-source autograder designed for advanced STEM courses, operating solely on standard on-premises equipment to maintain FERPA compliance without relying on third-party APIs. Tailored for LaTeX workflows prevalent in physics and engineering, LaTA employs a four-step process: ingest, segment, grade, and report, utilizing a locally hosted open-weight chain-of-thought LLM grader (gpt-oss:120b). It evaluates student submissions against a reference solution created by instructors and implements a YAML rubric for binary scoring on each item. LaTA was implemented in Winter 2026 for ME 373 (Mechanical Engineering Methods) at Oregon State University, assessing weekly assignments for around 60 students while ensuring data privacy by retaining all student information on institutional servers.
Key facts
- LaTA is a drop-in, open-source autograder for upper-division STEM courses.
- It runs entirely on commodity on-premises hardware, ensuring FERPA compliance.
- LaTA assumes a LaTeX-native workflow already adopted by many engineering and physics courses.
- It implements a four-stage pipeline: ingest, segment, grade, report.
- Uses a locally hosted open-weight chain-of-thought LLM grader (gpt-oss:120b).
- Compares student work to an instructor-authored reference solution with a YAML rubric.
- Deployed in Winter 2026 in ME 373 at Oregon State University.
- Graded every weekly assignment for approximately 60 students.
Entities
Institutions
- Oregon State University
Locations
- Oregon State University
- Corvallis
- United States