AIEd Knowledge BaseAIEd KB
On this page

Concept

Intelligent Tutoring

created 2026-08-09 · updated 2026-09-25

AI tutoring / intelligent tutoring — the use of AI to provide personalized, adaptive, scalable instructional support: from classical Intelligent Tutoring Systems (ITS) that model student knowledge, adapt instruction, and scaffold Problem Solving, to conversational and agent-based tutors built on LLMs. Effectiveness hinges on pedagogical design (Scaffolding, feedback quality, autonomy balance) rather than the model alone — see Measuring Whether LLM Tutors Teach or Solve: A Diagnostic for Educational Impact and Socratic Method.

Questions to Consider

  • An AI tutor's effectiveness is said to hinge on pedagogical design — scaffolding, feedback quality, the balance of autonomy and guidance — rather than on the underlying model. If you had to pick, which design choice most determines whether a student actually learns?
  • Classical ITS are praised for precision and transparency but criticized for inflexibility, while LLM tutors are flexible but can hallucinate or bypass learning. Where do you think the right balance lies, and why?
  • Large-scale field evidence found students tried an AI tutor but rarely used it productively — engaging it in only a fraction of their mistake sessions, often with bare answers. Why might students get access to a powerful tutor yet not use it in ways that help them learn?
  • The page contrasts traditional ITS that model a learner and adapt instruction with open-ended conversational tutors. What is gained and lost when a tutor can hold a natural conversation but doesn't know exactly why it made a decision?
  • One study found AI created a 'productive slowdown' — more time per question — but only improved learning when embedded in a mastery workflow that made mistakes consequential. What does that suggest about structuring effort so it pays off?
  • If hint design can inadvertently enable students to bypass learning, what would a well-designed hint system look like that supports struggle without giving away the answer?

Introduction

AI tutoring encompasses the use of artificial intelligence — particularly large language models and structured Intelligent Tutoring Systems — to provide personalized, adaptive, and scalable instructional support to learners. AI tutors take many forms: conversational tutors that engage in Socratic dialogue, scaffolded feedback systems that guide problem-solving, adaptive learning platforms that personalize content sequencing, and agent-based tutors that maintain long-term learner models. The effectiveness of AI tutoring depends critically on pedagogical design choices — scaffolding, feedback quality, and the balance between autonomy and guidance — rather than on the underlying model alone.

Historically, Mishra et al. locate ITS within AIED's lineage from 1960s-70s expert systems and Anderson's ACT/ACT-R cognitive tutors, whose structured control contrasted with Papert's constructionism.

ITS in the learner-modeling family

Intelligent tutoring is the classic application-side member of the learner modeling and adaptive instruction family. Its canonical architecture — domain model, student model, and pedagogical model — is precisely the "model a learner, then adapt instruction" pipeline the family describes. ITS consume the learner representations produced by Knowledge Tracing and Cognitive Diagnosis to select problems and scaffold guidance, which is why tutoring is so tightly coupled to those modeling methods. Within the family, ITS sits alongside adaptive learning (the real-time adaptation mechanism) and personalized learning (the broader goal) as the platforms that turn learner models into instruction.

ITS vs. LLM-based tutoring

The emergence of LLMs has created a productive tension in the tutoring field. Traditional Intelligent Tutoring Systems (ITS) offer precision and transparency — you know exactly why the system made a particular decision — but lack flexibility. LLM tutors offer natural dialogue and broad knowledge but can hallucinate, over-scaffold, or bypass learning entirely. Modern research increasingly explores hybrid approaches that combine structured ITS components with LLM flexibility. Reddig, Arora & MacLellan (2025) demonstrate this concretely in the Apprentice Tutor College Algebra ITS: supplying GPT-4 the tutor's interface structure and Bayesian Knowledge Tracing skill estimates within the prompt lifted logical-error diagnosis from 40% to 81% on factoring (and error identification overall to 87.8%) and produced ~66% error-targeted hints — direct evidence that embedding an LLM in an ITS's structured framework grounds generation, curbs hallucinated diagnoses, and yields context-aware corrective feedback, even though roughly a third of the hints remained too general, incorrect, or answer-revealing.

Intelligent Tutoring Systems represent one of the oldest and most researched areas of AI in education. Unlike general-purpose LLM tutors, ITS traditionally use structured approaches: domain models (what to teach), student models (what the learner knows), and pedagogical models (how to teach). These components enable fine-grained tracking of student progress, misconception diagnosis, and adaptive sequencing.

Key ITS research

  • Micro-randomized trials of a tutoring platform: Harrison et al. (2026) ran a four-week multisite individually randomized evaluation of Medly in GCSE Biology, Chemistry and Physics, with 644 of 929 students completing post-testing; allocation to the platform produced Hedges' g = 0.33 (95% CI 0.18 to 0.48) against business-as-usual self-directed revision, with positive estimates in all three subjects and no evidence of differential impact by disadvantage (Pupil Premium g = 0.28 versus non-Pupil Premium g = 0.35).

  • Hint button research shows that traditional ITS hint design can inadvertently enable bypass strategies, calling for more sophisticated Scaffolding approaches.

  • DeepTutor: Towards Agentic Personalized Tutoring provides a fully Open Source agentic tutoring framework with citation-grounded tutoring and difficulty-calibrated question generation.

  • Interpretable Knowledge Tracing addresses the opacity problem by producing interpretable cognitive quantities from LLM logits.

  • Engagement and structure, not capability, are the binding constraints (large-scale field evidence): Khanmigo (Oreopoulos & Low 2026) — a two-year cluster RCT in 18 middle schools — found 96% of students tried the AI tutor but the median engaged it in only ~17% of mistake sessions (mostly bare answers or prompt clicks), so gains (~0.06–0.08 SD) matched practice without AI. NUMI (Oreopoulos et al. 2026) found AI created a "productive slowdown" — more time per question and improved post-mistake recovery — but only reliably improved delayed learning when embedded in a mastery workflow that made mistakes consequential. The lesson: AI tutoring's value depends on getting students to use it productively and structuring it so effort pays off, more than on raw model capability.(Virtual Tutoring with Computer-Assisted Learning: An Experiment in Take-Up and Learning)

  • Outcome-based knowledge tracing for OBE tutors. Pradeesh et al. (2026) trace student mastery of Outcome-Based-Education course outcomes with a recurrent model that replaces learned concept relations with curriculum-validated OBE "affinity mappings" between course and program outcomes and adds a Memory Augmented Neural Network for cross-outcome impact. It reached 89.81% AUC on live engineering-program data, beating DKT, DKVMN, EKT, and SimpleKT baselines — a mastery signal aligned to a program's own outcomes that can drive adaptive problem selection in outcome-based tutoring.

  • Validate mastery models on future sessions, not retrospective fit. The mastery estimates an ITS uses to select problems may not generalize across time: Schuetze, Yan, and Carvalho (2025) found that BKT, BKT-with-Forgetting, and AFM reproduce learning trends only when fit retroactively to all sessions of a successive-relearning dataset — but under time-based cross-validation (training on one session to predict the next, the actual tutoring setting) they overestimate performance by 47–58%, fail to capture the spacing effect, and can mis-order practice conditions. Tutors relying on such tracers should be evaluated walk-forward and consider retention interval and between-session forgetting rather than assuming mastery persists.

  • Diagnosing mistakes on unstructured, handwritten-solution problems. Arthur (Yin et al. 2026) extends ITS-style feedback to Calculated Formula Questions in an Engineering Economics course, a domain where pen-and-paper solutions lack the structured digital data that previous AI tutoring work relied on. Rather than a general-purpose LLM, each question gets its own dedicated XGBoost "solution diagnosis backbone" — trained on curated, random-masking-augmented graded submissions to predict instructor rubric mistake labels from students' submitted numbers alone (precision 0.81, recall 0.79) — paired with a dialogue-based interaction that iteratively requests intermediate answers only when diagnosis confidence drops below a 0.8 threshold. It is a concrete case of the knowledge base's principle that a knowledge-grounded classifier can handle solution diagnosis even where full written solutions are unavailable.

  • Turn-level adaptive tutoring with a background evaluator. The MeduAI-SP tutor agent (Yang et al., 2026) exemplifies turn-level adaptive tutoring: a background evaluator monitored each student utterance against a 30-item OLDCARTS-linked history checklist plus communication criteria and triggered prompts only when warranted (24.1% of 4,815 messages). Corpus analysis exposed where such support is most needed — student talk was 62.2% history of present illness but only 4.3% empathic/supportive communication and 2.6% physical examination — suggesting ITS-style clinical tutors should target underused consultation components and late-encounter reasoning phases rather than symptom gathering alone, and that simulated-patient fidelity is high enough for outcome research (only ~0.68% of patient utterances showed clear fidelity problems).

  • Prompt-based micro-personalization of an LLM tutor (2026): Basu, Kakar & Goel (2026) personalize the Jill Watson Large Language Models (LLMs)/RAG (Retrieval-Augmented Generation) teaching assistant at the level of individual questions entirely through prompt engineering — conditioning responses on six learner dimensions (metacognitive Self-Assessment, abstraction, verbosity, perception, processing, understanding) and classifying each query's cognitive demand with a Bloom's Taxonomy classifier, yielding 96 learner profiles with no model retraining. NLP analysis of 2,910 responses and a human study found measurable, perceptible differences in response abstraction, verbosity, complexity, and processing style aligned with intended effects. It demonstrates response-form (rather than content) adaptation as a scalable, modular route to adaptive behavior in a deployed tutor.

Key ITS concepts

Historical context

The ITS field has produced landmark systems (Cognitive Tutors, Andes, AutoTutor) and continues to evolve. The Zerkouk et al. comprehensive ITS review catalogs this evolution. The tension between structured ITS and open-ended LLM tutoring is explored in the correct answer trap research and rethinking scaffolding for LLM tutors.

AI tutoring with LLMs: practical guidance

For instructors deploying AI tutors and developers building them, the knowledge base's findings translate into concrete practice:

Evaluate tutors on whether they teach, not just solve. A model that tops a solving leaderboard is not necessarily a good tutor — task-solving ability and learning-supportive behavior correlate only partially (r ≈ 0.42), and several models shift rank when scored on pedagogy. Report and scrutinize solving and pedagogy scores separately, and prioritize tutors that score on guiding questions, calibrated hints, and non-disclosive scaffolding over those that produce fast answers.(Measuring Whether LLM Tutors Teach or Solve: A Diagnostic for Educational Impact)(Methodologies for Improving the Quality of AI Tutoring in K-12 Education)

Design for pedagogical structure, not frequency. The educational payoff of AI tutoring depends on how the tool is used and designed, not on how often it is used. Instructor-designed tutors scoped to course objectives, learner proficiency, and a curated knowledge base outperform unstructured general-purpose chatbot use.(Instructor-Designed AI Tutors in University Foreign Language Education: A Mixed-Methods Study of Learner Motivation and Reflective Learning Experience Based on Self-Determination Theory) A preregistered randomized field experiment tested that claim against the realistic alternative rather than an AI-free baseline: 86 students in an online-MBA corporate-finance module were assigned to a tutor grounded in the module's own lectures, readings and problem sets or to a holdout that kept standard resources with consumer AI freely available, and tutored students gained 6.63 more points of 55 (95% CI [+1.95, +11.31], p = .007) while the share of short written answers at relational quality or better rose from 8% to 49% against 8% to 27%. Channel was the null margin — voice nearly doubled interaction density and cost 2.8× more to deliver, yet weekly mastery differed by −0.01 points of 11 (p = .98) — which makes delivery format an adoption and engagement lever rather than a learning technology. The holdout did not close the gap either: the nine students reporting external-AI use gained no more than the eight reporting none (+10.3 versus +12.8 of 55, p = .45).

Use iterative live evaluation to keep improving. Because LLMs are opaque, treat evaluation as the engine of improvement: instrument a small set of quality and engagement metrics, run live experiments on models, prompting, personalization, and agents, and let data drive changes — the same discipline Khan Academy applies to its K-12 tutor (Khanmigo).(Methodologies for Improving the Quality of AI Tutoring in K-12 Education)

Support the learner's autonomy, competence, and relatedness. AI tutors work best when they feel like a safe, structured practice space rather than an answer machine. Provide immediate, nonjudgmental Feedback; scope the tutor to the learner's level so competence is achievable; and preserve learner agency by keeping the tutor a complement to (not a substitute for) other instruction.(Instructor-Designed AI Tutors in University Foreign Language Education: A Mixed-Methods Study of Learner Motivation and Reflective Learning Experience Based on Self-Determination Theory)

Guard against answer disclosure. The central failure mode of LLM tutoring is giving the answer away, which inflates immediate performance while undermining durable learning. Use Socratic prompting, calibrated hints, and non-disclosive scaffolding — and measure outcomes on unassisted, transfer tasks, not just in-tool performance.(Measuring Whether LLM Tutors Teach or Solve: A Diagnostic for Educational Impact)(Socratic Method)

Separate diagnosis from feedback. LLM tutors reliably confirm correct steps but over-reject valid-but-suboptimal reasoning and over-validate incorrect solutions — and accurate diagnosis does not reliably yield actionable feedback.(Confirming Correct, Missing the Rest: LLM Tutoring Agents Struggle Where Feedback Matters Most) The coupling also cuts the other way: Reddig et al. (2025) found GPT-4 still crafted relevant, general-but-correct feedback ~74% of the time even after misdiagnosing an error (by restating the concept or expected answer format), yet almost all factually incorrect feedback followed a wrong diagnosis — so error identification remains the crux of actionable tutoring. A hybrid architecture works best: let a knowledge-grounded classifier handle solution diagnosis while the LLM focuses on open-ended scaffolding and dialogue.

AI tutoring as a spectrum of relational intensity

A unifying lens from the Stanford SCALE / NSSA brief (Turano et al. 2026) reframes AI tutoring as not a monolith but a spectrum defined by relational intensity — the depth and consistency of the human connection between student and tutor. The brief maps models from fully human-led tutoring (in-person or remote), through human-led with AI support (AI assists the tutor behind the scenes) and AI-led with human support (a human oversees and intervenes), to AI-only tutoring (no direct human oversight). The central finding: as direct human relationships decrease, the evidence base becomes thinner, and unresolved questions accumulate about student safety, developmental impact, and long-term efficacy.

  • AI augments, does not replace, high-impact tutoring. The brief's core message is that AI is best used to enhance tutor effectiveness and educator capacity within high-impact tutoring (regular school-day sessions, small-group ratios ≤ 1:4, well-trained consistent tutors, data-driven instruction, vetted materials, strong student-tutor relationships) — not to substitute for the human-led relationship that drives learning gains. This aligns with the knowledge base's broader finding that pedagogically designed AI tutors outperform general-purpose chatbots (Stanford evidence base, umbrella review).
  • Dosage, not model capability, is the binding constraint. The brief reports that AI-led tutoring inherits the dosage evidence base (~90 minutes weekly) only when scheduled, supervised, and protected within the school day; in a study of 181,000 students on a supplemental math platform, only 5% reached the recommended minutes and 41% never logged on, with teacher/school/district factors explaining 57% of usage variance. This converges with the field-experiment evidence above (Khanmigo, NUMI) that engagement and integration, not raw capability, determine whether AI tutoring helps.
  • Human oversight improves engagement and alignment but not dosage. The brief's RCTs found human check-ins raised elementary students' engagement with an AI platform but did not reach the dosage associated with gains nor improve reading achievement. The practical implication: schedule and supervise AI tutoring on the same terms as human-led tutoring, and design for enforcement of engagement — the opt-in requirement an on-device AI tutor reintroduces is exactly what live school-day tutoring eliminates.
  • The relationship is the irreducible element. The brief emphasizes that AI does not yet replicate human relationships, and that relationship-building (especially consistent tutor-student pairings) improves engagement, attendance, motivation, and outcomes. Developers describe using AI's strengths rather than replacing relationships — a design stance echoed in SafeTutors and affective tutoring research.
  • Safety and privacy are non-negotiable guardrails. For direct-to-student AI, the brief calls for strict student data privacy safeguards, Guardrails on student safety, and attention to the depth of unmonitored interaction — open questions remain about AI companions' effects on developing minds and prosocial development.

This relational-intensity framing is the "not a monolith" counterpoint to the field: evaluating any AI tutor should begin by asking which conditions of effective tutoring it reproduces, changes, or drops — rather than asking whether "AI tutoring works" in the abstract. The brief's human-led baseline can also be absent entirely: a six-week pre-test/post-test study across three Kwara State colleges of education gave 240 technical-education students a localized tutor fine-tuned on Nigerian technical curricula and reported roughly double the practical-skill gain of 120 conventionally taught peers. Practicals there ran above 100:1 student-to-lecturer ratios with feedback delayed by days, so the authors' rationale was not augmentation but equitable substitution for scarce instructional staff — casting the tutor as a capacity subsidy for rural colleges rather than a complement to plentiful human tutoring. Two design choices carry the practitioner lesson: the tutor explains an error rather than only flagging it, and a "Clarify" feature restates technical terms in Yoruba or Nupe, adapting to the learner's language context rather than only their knowledge state; hints rather than answers and a safe-to-fail space are credited with the reported confidence increase in most of the AI-assisted group against fewer than half of controls. The evidence is indicative only — no effect sizes, an unnamed base model, self-reported confidence — but it marks the boundary the brief leaves open, the setting where AI is asked to stand in for absent instructional capacity. It ties the tutoring page to K-12 policy (Educational AI Policy), Privacy, Pedagogical Safety, and the human-in-the-loop concerns throughout (Impact of real-time AI feedback on the technical skill acquisition of STEM students in Colleges of Education in Kwara State).

Practical design and development guidance

Design for learning, not just performance. The strongest causal finding is that unguarded AI tutors raise assisted practice performance but reduce unassisted learning — the performance–learning gap. Guardrail against answer-copying by scaffolding hints instead of answers (require a student attempt before revealing output), and verify gains on unassisted, closed-book measures rather than in-tool performance.(Generative AI without guardrails can harm learning: Evidence from high school mathematics)(Distinguishing performance gains from learning when using generative AI)

Make hints genuinely productive, not bypassable. Classic hint designs can enable "button-through" strategies that skip learning. Prefer hints that reveal reasoning steps incrementally (Socratic prompting) over hints that directly supply the next answer, and preserve productive struggle rather than over-helping.(Revisiting the Hint Button: Consistent Negative Associations Between Unproductive Hint Use and Learning Outcomes in Intelligent Tutoring Systems)(Rethinking Scaffolding in LLM Tutors: The Interactional Mismatch Between Benchmarks and Real-World Deployments)

Keep the human in the loop. Let teachers author or curate the problem sets and misconception prompts the tutor draws on, and surface the tutor's reasoning so its decisions are auditable. Interpretable Knowledge Tracing and explicit, external didactic layers make LLM tutor behavior traceable and reproducible.(Interpretable Knowledge Tracing)(A didactical-driven teacher assistant for a dimensional modeling course)

Model the learner, not just the dialogue. Attach structured Learner Modeling and Adaptive Instruction and Knowledge Tracing components to LLM dialogue so the system can adapt difficulty and diagnose misconceptions from evidence rather than responding fluently but blindly — quality depends on both the base model and how it is adapted.(EduClaw-Bench: A Long-Horizon Benchmark for Pedagogical LLM Agents with Simulated Learners)

Start from open tooling where possible. Open-source agentic tutoring frameworks (e.g. DeepTutor: Towards Agentic Personalized Tutoring) lower the barrier to a citation-grounded, difficulty-calibrated tutor you can inspect and extend.(DeepTutor: Towards Agentic Personalized Tutoring)

Effective tutoring requires continual adaptation: Zhang et al. (2026) evaluate whether LM tutors adapt to learners' evolving understanding at teacher-annotated decision points. They find frontier models default toward over-helpfulness and rarely push for rigor, and that evaluation-aware prompting improves but does not fully solve adaptivity. A systematic view of the RL-driven branch of this adaptation comes from Riedmann, Schaper & Lugrin (2025), whose review of 89 RL-in-education studies finds classical RL policies more consistently effective than Deep RL and RL adaptation delivering significant gains more often for guidance-related tasks (hints, Feedback) than for content scheduling — evidence that how an ITS adapts scaffolding matters as much as which algorithm it uses.

  • Deceptive overgeneralization (An et al. 2026) shows ITS mastery stopping rules (BKT, 95% threshold) can end practice before learners learn when to withhold a skill: learners who overgeneralized misapplied actions on first "do-not-act" items at 61.5%–100%, and targeted refrain-practice with constraint-naming Feedback reduced this to near-floor. Correctness-based mastery inference is necessary but not sufficient for ITS adaptivity.
  • Graph-based ITS for dynamic domains. A graph-based intelligent tutoring system combines an Evolving Knowledge Space Graph with generative AI content creation and Bayesian knowledge propagation — which showed the highest knowledge gains — supporting adaptive learning in dynamic curricula.
  • Rule-integrated LLM tutoring for procedural domains. Looi, Liu, and Sun (2026) tackle the inconsistency and pedagogical opacity of LLM tutors in primary mathematics through a design science study of a rule-guided system organized around a three-layer architecture — diagnosis → intent selection → constrained response generation. They formalize the distinction between rule-guided scaffolding (governed by auditable, replicable rules) and ad-hoc scaffolding (helpful moves difficult to audit or replicate). Evaluated via persona-based simulated dialogues and a classroom pilot with 40 Grade 5 students, rule-guided scaffolding improved interactional consistency, reduced premature answer-giving and early closure, and sustained cognitive engagement — while the classroom pilot surfaced interactional complexities, fragmented inputs, and attentional fluctuations that Simulation missed. This is a concrete blueprint for guardrailing Large Language Models (LLMs) tutors in well-defined procedural domains.
  • Multi-agent tutoring and automated assessment. Multi-agent tutoring systems are being benchmarked with synthetic, trace-based evaluation. ASTRA supports alone-tutor, pair-tutor, and pair-multiagent configurations with socially differentiated agents, enabling reproducible analysis of interaction and participation balance in introductory programming. In parallel, context-aware prompting automates coding of collaborative problem-solving skills from process data, supporting large-scale tutoring assessment.
  • Design with the learner's values, not just for learner performance. Value Sensitive Design work with community college students and instructors in developmental math shows that the ethical dimensions of ITS are not separable add-ons: engaging students and instructors directly produced 16 value-aligned features spanning explainability (comprehension-check interpretation, learning-path connection, communicating the model's confidence), learner control (control over re-assessment, review, pace, and AI-assistance level), and Privacy (data-repurposing control, permission prompts for sharing learning analytics and affective states). The study frames a persistent design tension — student agency vs. system-guided scaffolding — and notes that most institutions disable adaptive tutoring entirely, so value-informed design also depends on how (and whether) ITS AI capabilities are actually deployed.

Connected Concepts

Connected Articles

Connected FAQs

Metadata

DisciplineStem Education
LevelHigher Ed, K 12
Learning and instructionScaffolding
Technologies and techniquesAdaptive Learning, Intelligent Tutoring, Knowledge Tracing, Learner Modeling and Adaptive Instruction
Assessment and measurementFeedback
SourceHistoryEditChat

Embed this page

Copy the code below to embed a chromeless version of this page in a learning management system or other website. The embedded view hides the site header, navigation, and footer.