🧠 AI Ed Wiki

📄 Research Article

LaTA: A Drop-in, FERPA-Compliant Local-LLM Autograder for Upper-Division STEM Coursework

● high · 2026-05-15automated-grading, higher-ed, stem-education, llm, generative-ai, efficacy-study, feedback-loop, assessment-validity, formative-assessment, open-source

LaTA: A Drop-in, FERPA-Compliant Local-LLM Autograder for Upper-Division STEM Coursework Rodríguez (2026) — Oregon State University. Submitted to Computers & Education.

LaTA: A Drop-in, FERPA-Compliant Local-LLM Autograder for Upper-Division STEM Coursework

Summary

LaTA (LaTeX Teaching Assistant) is a privacy-preserving, drop-in autograder that addresses the core tension in educational LLM deployment: most Automated Grading systems send student work to third-party APIs, violating FERPA and exposing institutions to data risk.

System architecture:

  • Four-stage pipeline: Ingest → Segment → Grade → Report
  • Grading engine: gpt-oss:120b running entirely on local commodity hardware (single Mac Studio)
  • Rubric format: YAML with binary per-item scoring, comparing student work against instructor-authored reference solutions
  • Cost: $0 marginal cost per assignment, 1-3 minutes per submission

    • Real-world deployment (Winter 2026):

  • Course: ME 373 at Oregon State University (~200 students)
  • Grading error rate: 0.02-0.04% per rubric line item — near-perfect accuracy
  • Operational benefit: Enabled regrading of corrected assignments and expanded TA office hours

    • Learning outcomes (vs. traditional cohort):

  • Midterm exam: +11% performance gain
  • Final exam: +8% performance gain
  • Confidence gains: Δ ≥ +1.49 Likert points on every learning objective (N=159, p < 10⁻²⁷)

    • These results provide strong evidence for the Feedback Loop hypothesis: faster, more consistent feedback drives both learning and confidence. The deployment demonstrates that Generative AI grading can be both FERPA-compliant and pedagogically effective, addressing concerns raised in Assessment Validity and Formative Assessment discussions.

    LaTA's success connects to the broader STEM Education and Higher Ed landscape, showing that LLM-based grading can move beyond Short Answer Scoring Quality Degradation concerns when properly designed with instructor-authored rubrics and reference solutions. The open-source, zero-marginal-cost model aligns with Principled AI Education principles.

    Connected Concepts

  • Automated Grading
  • Feedback Loop
  • Generative AI
  • Assessment Validity
  • Formative Assessment
  • STEM Education
  • Higher Ed
  • LLM

    • Connected Articles

  • Short Answer Scoring Quality Degradation
  • Principled AI Education

    • Citation

    Rodríguez, J. A. (2026). LaTA: A drop-in, FERPA-compliant local-LLM autograder for upper-division STEM coursework. arXiv:2605.05410. Submitted to Computers & Education.