TopThi — An Adaptive Learning Loop for National High-School and Aptitude Exam Prep
TopThi turns every attempt into diagnostic data, then uses learning science to decide what each student should practice next. It is XO Edu Lab's living lab, running in production at topthi.vn.
The Problem
Doing more practice tests does not guarantee progress. Real personalization means knowing which skill a student is weak in, targeting practice there, and proving that score gains are real.
A total score doesn't show where a student is weak
A 6.5 doesn't tell a student what to practice next. Ability has to be measured per skill.
Random practice creates an illusion of progress
Redoing familiar tests pushes practice scores up but drifts into memorizing answers. Progress has to be measured on new questions.
A large question bank with no structure
Hundreds of thousands of questions with no skill tags, skewed difficulty labels and some wrong answer keys. Without clean data there is no personalization.
Analysis must never slow down submission
All ability computation runs in background queues. Students get results instantly; their skill profile updates right after.
Feature Map
Question Bank Engine
- Skill map by Subject · Grade, aligned with the 2018 national curriculum and aptitude exams
- AI skill tagging with human review
- Difficulty calibrated from real attempt data
- Automatic detection of suspected wrong answer keys, with student error reports and AI re-checks
- Suspect questions are automatically excluded from generated sets
Exam Engine
- Fixed exams and per-student generated sets
- All 2018-curriculum question formats: multiple choice, multi-part true/false, short answer, grouped reading
- Instant automatic grading on submit
- Server-verified exam timing, with an answer snapshot
Adaptive Learning Engine
- Elo/IRT ability model per student and per skill
- Placement test weighted by exam blueprint
- Daily interleaved practice at the right difficulty
- Prerequisite skill graph: switches to foundations when they are weak
- SM-2 spaced review scheduled from attempt results
Learning Analytics
- Predicted score with a confidence interval and the 3 topics most worth investing in
- Per-skill ability profile for students and admins
- Learning-outcomes report: pre/post periodic assessments, predicted-score error
- D1/D7/D30 retention, exam traffic sources
The TopThi Learning Loop
This is a loop: step 5 feeds back into step 2 (Diagnose), and tomorrow's practice set is recomputed from there.
The Science Behind It
Every TopThi design decision rests on a learning mechanism backed by research. Main sources: Make It Stick (Brown, Roediger, McDaniel, 2014) and How People Learn I & II (National Academies, 2000, 2018).
Testing effect
Retrieving from memory makes learning stick better than rereading. Getting it wrong and then receiving feedback is also learning.
Students practice with short daily sets instead of rereading material. Every wrong answer becomes diagnostic data.
Spaced practice
Reviewing at intervals, just before forgetting, produces more durable memory than cramming.
An SM-2 review schedule for each skill of each student.
Interleaving
Mixing problem types forces learners to identify which method applies, just like a real exam.
Each practice set mixes 3–4 topics, with at least 2 questions per topic.
Desirable difficulty & zone of proximal development
Work should be hard enough to demand effort, yet still achievable.
Question difficulty targets roughly 70% correct. When foundations are weak, the system switches to prerequisite skills.
Formative assessment & feedback
Feedback works when it is specific, timely and points to the next step.
Students get a per-skill report, see the misconception under a wrong option, and know the 3 topics most worth their time.
Measurement & verification
Ability and question difficulty sit on the same scale. Progress must be measured on new questions.
An Elo/IRT model with guessing correction, plus periodic assessments using questions the student has never seen.
TopThi optimizes for real progress, not the feeling of progress. Effective practice often feels harder and slower.
A personal practice set every day
Each time a student opens the Practice page, today's set is recomputed from their full attempt history. There is no pre-written 30-day plan.
- 15 questions, 3–4 interleaved topics
- Predicted score per subject
- Prioritizes skills due for review
Technical Notes
Review: know why you got it wrong
Every wrong option is linked to a typical misconception. A student who picks it sees the explanation right under the option they chose.
- Per-option misconception notes
- Report a question error
- Verified explanations
Technical Notes
Measuring progress on new questions
Periodic assessments draw from a separate question bank and exclude every question the student has seen, so score gains are not just familiarity.
- Assessment every 7 days
- Ability per skill
- Outcomes report for admins
Technical Notes
System Architecture
Backend
Frontend
Admin
Data & infrastructure
Integrations
"Our first model added mastery points for every question attempted, so a student guessing at random could still be rated "good". When we switched to an Elo/IRT model, many topics turned out to have been overrated. The lesson: without a skill map and clean data, AI cannot personalize anything — and that is where most of the effort goes."— The XO Edu Lab Team