Adaptive learning paths

A fixed course sequence bores strong learners and loses weak ones. Adaptive learning systems keep a model of each learner's knowledge, choose the next exercise or lesson that fits it, and tell teachers where a group is stuck. Classic knowledge tracing now combines with LLMs that generate, explain and tag content. In the EU the decisive question is whether the system merely recommends practice or evaluates learning outcomes and steers the learning process, which can make it high-risk under the AI Act.

Problem & value

• Today: one pace for the whole class or cohort; differentiation is manual and limited by teacher time. Corporate training is mostly a linear course catalogue with a completion checkbox.
• AI value: diagnose prior knowledge, skip what is mastered, repeat what is not, space repetition, and give teachers a live map of gaps so they can intervene where it counts.
• Evidence: the field has a long research base (intelligent tutoring systems, mastery learning, knowledge tracing). Reported effects vary strongly by subject, dosage and implementation; vendors publish efficacy studies but results should be checked against independent evaluations (e.g. Evidence for ESSA, EEF toolkit) rather than marketing figures [1][2].

Variants

• School (K-12): maths, languages and reading practice platforms with teacher dashboards (e.g. Bettermarks, Anton, Khan Academy, Mathematics-focused tools).
• Higher education: adaptive courseware and prerequisite-aware study plans, personalised revision in large first-year courses.
• Corporate L&D / compliance training: skills-gap based learning paths, recommended courses from a skills ontology (see EDU-006).
• Vocational and exam preparation: adaptive question banks (driving licence, bar, certification), often with spaced repetition (EDU-016).
• Private learners: language and skill apps with adaptive difficulty (Duolingo, Babbel [EU], Memrise).

How it works

1. content + competency graph (skills, prerequisites, difficulty tags)
2. diagnostic / placement test
3. learner model (knowledge tracing: BKT / PFA / DKT / IRT-based mastery per skill)
4. next-activity policy (rules, bandit or recommender: zone of proximal development, spacing)
5. learner activity (exercise, text, video, LLM explanation or hint)
6. response logging
7. model update
8. teacher/trainer dashboard + overrides

• Learner model: Bayesian Knowledge Tracing, performance-factors models, item response theory (IRT), or deep knowledge tracing (transformer-based, e.g. SAKT/AKT). Simple models are often as good as deep ones and far easier to explain.
• Content tagging: LLMs map items to curriculum standards, estimate difficulty and prerequisites; embeddings plus a vector index find similar items and remediation material.
• Generative layer: LLMs write hints, alternative explanations and fresh practice items at a requested level (EDU-001, EDU-002, EDU-004). Facts and answers must be checked against curated content or a solver.
• Policy: rule-based sequencing for transparency, optionally contextual bandits to optimise engagement or learning gain.
• Human-in-the-loop: teachers set goals, can pin or skip content, review flagged gaps and decide on any grading or placement consequence; the system's mastery estimate is advice.

Data preparation

Adaptive learning needs three data sets: a competency framework with prerequisites, tagged content items (exercises, videos, texts with skill and difficulty labels), and learner interaction logs (answers, hints, time on task). In an SMB or training department these live in the LMS, in SharePoint or file shares, in slide decks and PDFs, and often only in trainers' heads.

Typical gaps: content without skill or difficulty tags, scanned handouts, duplicate versions of the same exercise, no explicit prerequisite structure, too little interaction history for a cold start, and unclear access rights to learner data.

Preparation steps: (1) define the competency model with subject experts and build a category scheme (PREP-019); (2) tag content with skill, difficulty and format using an LLM draft that experts review (PREP-010); (3) digitise paper material (PREP-006); (4) extract logs from the LMS via xAPI or export (PREP-005); (5) assemble a diagnostic pre-test and a gold-standard item set (PREP-018); (6) pseudonymise learner data for model development (PREP-022). Tools: LMS exports, Moodle/H5P, a spreadsheet for the competency map, an LLM for tagging drafts.

Rough share of effort: about 30-40 percent, dominated by expert time for the competency model and content tagging. See also GUIDE-003.

Implementation paths

Which path for whom

• Enterprise: Path A, because adaptive engines in established LMS/LXP products are mature and corporate training is a commodity function.
• Public (schools, universities): Path C, because learner data of minors and students is sensitive and open-source stacks (Moodle, Open edX, H5P) on sovereign EU hosting fit procurement rules.
• Developer: Path B, because a custom mastery model plus LLM explanations via EU-hosted APIs is the fastest way to build a differentiated product.

Path A — Buy (recommended for corporate training departments)

• K-12 / education: Bettermarks (DE, maths, EU), Anton (DE, EU), Khan Academy with Khanmigo, IXL, Century Tech (UK), Smart Sparrow/Pearson courseware, Squirrel AI (China-based), Knewton successors in Wiley and others. Check data residency and DPA.
• LMS-integrated: Moodle (open source, EU-origin) with adaptive lesson/quiz plugins, Canvas, Blackboard Ultra, itslearning (EU), Microsoft Learning Accelerators, Google Classroom.
• Corporate: Degreed, Docebo, Cornerstone, SAP SuccessFactors Learning, Coursera for Business, Sana (EU, Sweden) and Learnship-style platforms with skills-based recommendations. Skills ontology quality is the differentiator.
• Choose A when: curriculum is standard, there is no need to own the learner model, and a vendor has a signed GDPR DPA and an EU-hosted offering.

Path B — Build on cloud APIs

• Models: frontier LLMs through EU regions or EU providers (Mistral, Azure OpenAI EU Data Zone, Google Vertex EU, AWS Bedrock Frankfurt) for content tagging, hints and explanations; embeddings for similarity.
• Learner model: own service implementing BKT/IRT or a recommender with open libraries (pyBKT, py-irt, scikit-learn, Implicit/LightFM); interoperability through IMS standards (LTI 1.3, Caliper analytics, xAPI) and competency frameworks (ESCO, CASE) [3][4].
• Integration: LTI 1.3 into the LMS, SSO via school identity (e.g. Germany's Verbund identity services, Microsoft Entra), event pipeline into a warehouse.
• Choose B when: you have a distinctive curriculum or publisher content library.

Path C — Self-host (open source; recommended for schools and universities)

• Platforms: Moodle or Open edX with adaptive modules, OATutor (open-source adaptive tutor), Kolibri (offline-first, Learning Equality) for low-connectivity settings.
• Models: open-weight LLMs (Mistral Small/Large, Qwen3, Llama, Gemma) via vLLM or Ollama for hints and tagging; BGE-M3 or multilingual-e5 embeddings; Qdrant or pgvector.
• Hardware: the learner model itself is light (CPU, HW1); add one 24-32 GB GPU (HW2) only for local LLM hints. Strong fit for school authorities needing on-premise data.

Path D — Custom training

Worth it for large platforms with millions of interactions: train deep knowledge-tracing or item-difficulty models on your own logs, or fine-tune a small LLM on curated hints. Needs labelled content metadata and many responses per item (typically thousands) for calibration; cold start for new items stays a problem.

State of the art (2026)

• Knowledge tracing: transformer and graph-based models improve prediction benchmarks (ASSISTments, EdNet), but gains over BKT/IRT in real learning outcomes are modest and models are harder to explain.
• LLM-enabled content: on-the-fly item generation, rewriting at reading level and language, and tutoring dialogue (see EDU-004) have made content supply the cheap part; validity of generated items and alignment with the curriculum is now the bottleneck.
• Standards: LTI 1.3, xAPI/Caliper and open skills taxonomies (ESCO) ease portability.
• Limits and failure modes: cold start for new learners and items; wrongly tagged prerequisites; gaming the system (hint abuse, guessing); filter bubbles that narrow the curriculum; optimising engagement instead of learning; bias when the model is calibrated on unrepresentative populations; LLM-generated errors. Engagement time is a poor proxy for learning gain.

Shared resources & combinations

• Same stack: EDU-002 question generation (item bank and difficulty), EDU-003 feedback, EDU-004 tutoring (generative layer), EDU-016 spaced repetition (scheduling), EDU-001 content creation.
• Downstream: EDU-007 learning analytics and early warning (same event data, different and more sensitive purpose); EDU-006 corporate microlearning; EDU-009 teacher differentiation.
• Shared infrastructure: vector database and embedding service, event pipeline, one LLM gateway with guardrails serving several education use cases.

Effort, cost & KPIs

• Path A: weeks to pilot; licences per pupil or per seat; teacher training is the main cost.
• Path B: 3-6 months for a robust learner model, content tagging and LMS integration.
• Path C/D: 6+ months; content calibration and ongoing curation dominate cost.
• KPIs: learning gain (pre/post, ideally with a comparison group), time to mastery, retention after 4-8 weeks, completion rate, share of recommendations accepted or overridden by teachers, equity gaps between learner groups, content error reports.

People & change

Involve management early for the decision and budget, the department lead (head of training or the teaching staff) as process owner, and a pilot group of learners and trainers from the start. IT handles LMS integration and accounts; the DPO is needed before any pilot because learner data is personal and, for pupils, concerns minors; a DPIA is likely (systematic evaluation of people, GUIDE-005). HR and the works council (or staff council in the public sector) must be involved because the system logs individual learning behaviour and is objectively suited to monitor performance: co-determination under §87(1) no. 6 BetrVG applies, and §90(1) no. 3 requires early information. Agree in a works agreement that learning data is not used for appraisals (GUIDE-004).

Support AI literacy of trainers and teachers through short workshops (Art. 4 AI Act, GUIDE-006). Typical concerns: "the machine labels me" and loss of teacher autonomy. Address them with transparent explanations of why an item was recommended, learner opt-outs of non-essential tracking, and teachers who can override every recommendation. Start with one course, measure, then scale (GUIDE-008).

Risks, regulation & compliance

• EU AI Act: Annex III point 3 lists as high-risk AI systems intended to evaluate learning outcomes, including when those outcomes are used to steer the learning process, and to determine access to or assess the appropriate level of education [5]. A pure content recommender that does not assess learners may be minimal or limited risk; a system whose mastery score drives placement, tracking, grading or progression is likely high-risk (risk management, data governance, human oversight, logging, conformity assessment). Obligations for Annex III systems were postponed from Aug 2026 to 2 Dec 2027 by the Digital Omnibus on AI (adopted mid-2026; verify against the Official Journal text) [5][6].
• Art. 5 prohibition: emotion recognition in educational institutions (e.g. inferring frustration or attention from webcam or voice to adapt content) is prohibited, except for medical or safety reasons. Do not use affect sensing to drive adaptation [5].
• Transparency: chatbots and synthetic content must be disclosed (Art. 50); AI literacy duty (Art. 4) for teachers and staff.
• GDPR: pupil data is personal data, often of children (Art. 8 consent age, strengthened protection); DPIA usually needed for systematic profiling; purpose limitation against reuse for ranking or marketing; German state school laws and DSK guidance on schools apply. Prefer EU hosting and a data processing agreement; no training of vendor models on pupil data [7].
• Pedagogical risks: over-reliance on automation, narrowed curriculum, stigmatising levels. Mitigations: teacher override, explainable recommendations, regular bias and equity audits, learners can see and contest their model.

Getting started

1. Define the learning objective and decide whether the system only recommends practice or influences grades and placement (this sets the AI Act class).
2. Map the curriculum into skills and prerequisites; start with one subject and grade.
3. Pilot a Path A tool or Moodle plugin with two classes and a comparison group, measuring pre/post learning gain.
4. Run the DPIA, sign the DPA, and agree on teacher override and parent information.
5. Add an LLM layer for hints and explanations only after the learner model is validated.

Sources

1. Evidence for ESSA, Johns Hopkins — ratings of education programmes — https://www.evidenceforessa.org/
2. Education Endowment Foundation — Teaching and Learning Toolkit — https://educationendowmentfoundation.org.uk/education-evidence/teaching-learning-toolkit
3. 1EdTech — LTI 1.3 and Caliper Analytics — https://www.1edtech.org/standards/lti
4. ESCO — European Skills, Competences, Qualifications and Occupations — https://esco.ec.europa.eu/
5. Regulation (EU) 2024/1689 (AI Act), Annex III and Art. 5 — https://eur-lex.europa.eu/eli/reg/2024/1689/oj
6. European Commission — AI Act implementation and Digital Omnibus — https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai
7. European Commission — Ethical guidelines on the use of AI and data in teaching and learning for educators — https://op.europa.eu/en/publication-detail/-/publication/d81a0d54-5348-11ed-92ed-01aa75ed71a1
8. pyBKT — Bayesian Knowledge Tracing library — https://github.com/CAHLR/pyBKT
9. OATutor — open adaptive tutoring system — https://github.com/CAHLR/OATutor