Skip to content
Studeia Docs

Adaptive learning: Bayesian mastery & misconception lifecycle

Studeia tracks per-concept mastery with a Bayesian model, runs a 3-state misconception lifecycle, schedules spaced repetition (SM-2), and adapts the tutor's pedagogical strategy to each learner.

2026-05-31 7 min
Resposta curta

Studeia personalizes learning with a Bayesian per-concept mastery model (ConceptMastery, Beta distribution), a 3-state misconception lifecycle (active → resolving → resolved) with full traceability and evidence, episodic memory of what worked, and spaced repetition (modified SM-2) over a prerequisite knowledge graph. The PedagogicalAgent adapts the tutor's strategy to each mastery level, and inline quizzes in the chat verify and resolve misconceptions in real time.

The student model

Before each tutor turn, StudentModelService assembles an enriched snapshot:

  • ConceptMastery — a Bayesian probability per concept (Beta distribution) with a confidence interval and a relative learning speed.
  • Misconceptions — active/resolving items with an evidence trail.
  • Episodic memory — what worked before (an analogy that landed, a breakthrough), scored by effectiveness.
  • Quiz context — average score, pass rate, weak areas.

Pedagogical strategy by mastery

MasteryStrategy
< 0.3direct instruction (clear explanation, concrete examples)
0.3–0.5scaffolding (progressive hints)
0.5–0.7socratic (guiding questions)
0.7–0.9guided practice (exercises + feedback)
> 0.9challenge (complex problems, connections)

Thresholds are configurable per tenant (PedagogicalConfig), with overrides by domain, age and learning style. The agent also adjusts when chat mastery and quiz performance diverge (shallow understanding vs. a quiet but capable student).

Misconception lifecycle

Misconceptions move through active → resolving → resolved with an evidence trail. They are created with full traceability — source (chat / quiz / interactive / inline_quiz), the triggering question, the student's answer, a context snippet, and linked concepts. The tutor injects active misconceptions into its prompt and addresses them proactively; frequent ones trigger an inline quiz in the chat. Resolution requires mastery + quiz/interactive evidence; re-detection regresses to active. Teachers can also resolve/reopen manually.

Spaced repetition & knowledge graph

A modified SM-2 algorithm schedules reviews per topic, with difficulty calibration from correctness and a per-subject prerequisite knowledge graph so reviews and new concepts arrive in the right order.

Inline quizzes

When the tutor detects a frequent misconception it can render an interactive inline quiz in the chat. Correct answers are kept server-side (Redis, never sent to the client before submit); on submit, mastery updates and the misconception can transition toward resolved — and the tutor follows up proactively.

See also

FAQ

How does Studeia personalize learning?

It builds an enriched student model: ConceptMastery (a Bayesian Beta-distribution probability per concept), active/resolving misconceptions with an evidence trail, episodic memory of what worked, and quiz performance. The PedagogicalAgent then picks a strategy per mastery level — direct instruction, scaffolding, socratic, guided practice or challenge.

What is the misconception lifecycle?

Each detected misconception moves through three states: active → resolving → resolved, with an evidence trail. It is created with full traceability (source chat/quiz/interactive, the question, the student's answer, linked concepts), the tutor addresses it proactively, and it only resolves when mastery and quiz/interactive evidence confirm it. Re-detection regresses it to active.

Does Studeia do spaced repetition?

Yes. A modified SM-2 algorithm schedules reviews per topic, with difficulty calibration based on correctness and a knowledge graph of prerequisites per subject, so reviews land before the learner forgets.

Veja tambem

Adaptive learning: Bayesian mastery & misconception lifecycle