Vietnamese LLM for Education: Teaching Aids and Grading

01/10/2026

My literature teacher used a red pen, and at the bottom of every essay she wrote a few lines by hand. Looking back, I realize how many evening hours those few lines cost her. Forty students, one essay a week. That is a reasonable starting point for talking about a Vietnamese LLM in education: the tool is not there to replace her, but to hand back some of the time that repetitive work takes away.

Teaching assistant: answers when the teacher is not there

Students get stuck at ten at night, halfway through homework, with no one to ask. A teaching assistant built on an LLM and loaded with the actual textbook and syllabus of a course can re-explain a concept, give another example, and ask questions back to check understanding. Design decides quality. If the model just hands over answers, the student copies and learns nothing. If it hints step by step and asks guiding questions, it is closer to what a good tutor does.

Vietnamese adds its own concerns. The language of instruction in primary school differs sharply from university: forms of address, sentence length, everyday examples. A model that answers a fourth grader in the tone of a university textbook is aimed at the wrong audience even if the content is correct. System prompts and suitable data handle much of this, and it is one place where finetuning an LLM genuinely helps.

Grading: support, not replacement

Machines have graded multiple choice for ages. The hard part is open-ended work: paragraphs, worked math solutions, essays. An LLM can read a submission, compare it against a rubric, point out reasoning errors, spelling mistakes and passages lacking evidence, and draft comments. The teacher reviews, edits, and decides the grade.

I want to be clear about limits here, because this is where misuse is easiest.

Scores have consequences: an LLM can grade inconsistently, giving two different results for the same essay run twice. Final grades should be set by the teacher.
Style bias: models tend to favor smooth, long, polished prose. A student who writes briefly but sharply may be marked down.
Originality: an essay with an unusual angle is sometimes treated as off-topic, because the model learned from average writing.
Explaining mistakes: the machine's comments can sound convincing and be wrong, especially in math and multi-step reasoning.

The practical use is to let the machine do a first pass and comment on surface errors, while the teacher focuses on depth and on the students who need attention.

Personalized learning paths

This is the most attractive part on paper and the easiest to oversell. The idea is simple: from each student's work, questions and mistakes, the system sees where they are weak, suggests suitable practice and adjusts difficulty. A student who never got fractions would review fractions before moving to equations, instead of being swept along by the class schedule.

I am careful on this topic for several reasons. Diagnosis quality depends on data, and learning data is usually sparse and noisy. "Personalization" can easily become locking a student into a narrow path chosen by a prediction. And student data, especially from minors, needs strict protection: where it is stored, who sees it, for how long, and whether parents know.

What teachers should keep

In my view some things should not go to a machine: judging a student as a whole person, deciding final rankings, and noticing that a child is going through something at home. My old literature teacher could tell from handwriting which student was off that day. An LLM cannot do that, and I am not sure it should.

On our side

AIVISION is an AI company in Vietnam that trains and finetunes LLMs for Vietnamese and for specific domains. We have released L1.0, an LLM for Vietnamese, and education is one application direction we care about. I am not citing specific results because we have no data to assert them. I am describing how we think it should be designed: grounded in real curricula, with tone suited to age, and always with a teacher in the loop.

If you work at a school or learning center and want to talk further, information about the models is on AIVISION.