Intelligent Tutoring Systems: A Comprehensive Historical Survey with Recent Developments

1 INTRODUCTION

• The field’s early educational-computing work began with CAI in the 1970s, when Carbonel proposed adding AI to address its limitations.The paper describes ITS design as an engineering field and notes its growth over the following decades [5].Source: §1 · p. 1
• Human tutoring’s reported two-standard-deviation learning advantage helped motivate AI researchers to develop individualized computer tutors.Bloom’s 1984 study compared tutoring with traditional group teaching.Source: §1 · p. 1
• ITS use multimedia to present information and have been used in homes, schools, universities, businesses, and governments.One stated goal is to better understand student behavior through interaction.Source: §1 · pp. 1–2
• ITS use AI to model what to teach, whom to teach, and how, while conventional CAI generally cannot monitor solution steps or provide instant help.ITS research also draws on computer science, cognitive psychology, and educational research, with differing goals and terminology across these disciplines.Source: §1 · p. 2

2 RELATED SURVEY PAPERS

• The paper groups prior ITS surveys into two categories: general discussions of the field and reviews focused on a specific ITS dimension.Source: §2 · p. 2
• General surveys covered ITS components and evolution, history, evaluation, and future directions, while four tutor case studies examined student interaction and development issues.Nwana’s 1990 survey discussed Computer-Assisted Instruction and popular ITS of its era.Source: §2 · p. 2
• Topic-focused surveys examined authoring tools, conversational ITS [13], constraint-based tutors, and tutor behavior in general and ill-defined domains [15].The authoring-tools work characterized individual tools, while Mitrovic reviewed the history and advanced features of constraint-based tutors.Source: §2 · pp. 2–3

3 WHY THIS SURVEY

• Beginners face difficulties understanding ITS operation, AI technologies, learning theories, system types and interaction, and educational effectiveness.Source: §3 · p. 3
• The paper’s goal is an updated, in-depth, demystifying introduction that presents core ITS aspects and directs readers to older and recent references.Source: §3 · p. 3
• The survey is not intended to focus on one ITS topic or provide exhaustive detail; readers can use it as a starting point before consulting more detailed sources such as Woolf.Source: §3 · p. 3

4 HUMAN TUTORS VS. COMPUTER TUTORS

• One-on-one human tutors guide learners through difficulties with hints, feedback, and strategies rather than simply supplying solutions [18][2][19].The passage also notes that self-directed learning can build knowledge and reasoning but may take more time and effort.Source: §4 · p. 3
• Model tracing matches a student’s steps against a rule-based domain expert and provides explanatory feedback, hints, or suggestions after errors; otherwise, it silently tracks progress.Source: §4 · p. 3
• Experiments cited in the paper report that model-tracing tutors support learning in domains including geometry and LISP [20].Source: §4 · p. 3

• Human tutoring had effect size 0.79 versus no tutoring, while step-based tutoring (0.76) was nearly as effective.Substep-based tutoring was 0.40 as effective as no tutoring.Source: §5 · p. 4
• ITS had no significant effect on K–12 math outcomes over short periods, but effectiveness appeared greater after a full school year or longer and among general students than low achievers.Source: §5 · p. 4
• ITS was associated with greater achievement than teacher-led and large-group instruction (effect size 0.42), non-ITS computer instruction (0.57), and textbooks or workbooks (0.35).Source: §5 · p. 4
• ITS showed no considerable difference from individualized human tutoring (-0.11) or small-group instruction (0.05), and effects varied with ITS features, student attributes, and domain knowledge.Source: §5 · p. 4
• 92% of 50 studies favored ITS over conventional classes; 39 reported gains up to 0.66 median effect sizes, while standardized-test effects were 0.13.Source: §5 · p. 4

• The typical ITS has domain, student, tutor, and user-interface models for content, learner state, pedagogical decisions, and access to domain knowledge [8][24].Source: §6 · p. 5
• The outer loop selects the next task using student history and background; the inner loop monitors solution steps and provides interventions such as hints or feedback.Source: §6 · p. 5

6.1 Domain Model

• The domain model stores expert-derived facts, concepts, strategies, and rules; it supplies explanations and solutions, evaluates learner performance, and depends on suitable knowledge representations.Source: §6.1 · pp. 5–6

6.1.1 Cognitive Model.

• Cognitive tutors use rule-based task models and common-error libraries to give stepwise feedback across problem-solving strategies [34].The figure illustrates production and buggy rules for computing an angle; the model can represent multiple paths to a correct answer.Source: §6.1.1 · p. 6, Figure 3 · p. 7
• ACT-R represents declarative knowledge as chunks and procedural knowledge as IF-THEN productions; model and knowledge tracing compare steps and track acquired skills [35][36].Cognitive Mastery Learning estimates the probability of knowing each skill using four parameters.Source: §6.1.1 · pp. 6–7
• Building complete model-tracing tutors requires multiple software components and a professional team, creating high cost and time demands that limit practical use.Source: §6.1.1 · p. 7

6.1.2 Constraint based Model (CBM).

• CBM checks whether solutions violate domain constraints rather than enumerating solution paths, accepting unmodeled valid strategies and supporting some ill-defined tasks [42].It focuses on principles shared by correct solutions, avoiding the extensive bug libraries model tracing may require.Source: §6.1.2 · pp. 7–8

6.1.3 Expert Approach.

• Expert-system approaches can use rules, neural networks, decision trees, or case-based reasoning, and can represent broader domains than cognitive or constraint-based models [44].Source: §6.1.3 · p. 8
• Fournier-Viger et al. describe comparing generated expert solutions with learner work or comparing ideal solutions with learner work; some systems are costly and provide unsuitable explanations.Source: §6.1.3 · p. 8

6.2 Student Model

• A student model represents learner knowledge and skills dynamically, using inferred mastery and behavior to guide personalized feedback, misconception correction, and instructional adaptation.Design considers what the student knows, the knowledge needed for a problem, their understanding of the mechanism, and their solution strategies.Source: §6.2 · p. 9

6.2.1 Overlay Model.

• The overlay model treats student knowledge as a subset of domain knowledge and marks gaps as deficits, but omits misconceptions, cognitive needs, preferences, and learning styles [51].Implementations range from Boolean mastery indicators to qualitative levels; systems often combine overlays with other approaches.Source: §6.2.1 · pp. 10–11

6.2.2 Stereotypes Model.

• Stereotypes initialize learner models by assigning users to groups with shared features, reducing individual knowledge elicitation, but group assignments can be inaccurate and costly to construct.Systems may combine stereotypes with fuzzy logic or machine learning, as in INSPIRE and Web-PVT [63].Source: §6.2.2 · pp. 11–12

6.2.3 Perturbation Model.

• Perturbation models extend overlays by representing misconceptions; discriminating problems can help identify wrong rules, while bug libraries are costly to build and maintain [67][49].Bug libraries may enumerate observed errors or list common misconceptions generatively.Source: §6.2.3 · p. 12

6.2.4 Constraint Based Model.

• Constraint-based modeling diagnoses each solution step by checking constraints and identifies errors and violated principles without requiring a runnable expert module or extensive bug studies [70].It has been applied across tutoring domains.Source: §6.2.4 · p. 12

6.2.5 Cognitive Theories.

• Cognitive theories support student modeling and error diagnosis; HPR models recurring inference patterns, while MADM selects among alternatives with multiple, often conflicting attributes [76].Examples include HPR-based systems RESCUER and F-SMILE, and Web-IT’s use of MADM with an age stereotype [78][79].Source: §6.2.5 · p. 13

6.2.6 Bayesian Networks.

• Bayesian networks represent student knowledge, misconceptions, emotions, and other factors as probabilistic relationships, supporting assessment, prediction, diagnosis, and adaptive feedback [49].ITS applications include Andes for long-term knowledge assessment and plan recognition, and systems modeling emotion or cognitive states [84][88][90].Source: §6.2.6 · pp. 13–14

6.2.7 Fuzzy student modeling.

• Fuzzy logic represents uncertain student characteristics with qualitative values rather than binary choices, and an evaluation in programming instruction reported increased satisfaction and performance, improved adaptivity, and more reliable decisions [92][94].Fuzzy membership values range from 0 to 1.Source: §6.2.7 · p. 14

6.3 Tutor Model

• The pedagogical module uses student traits to choose what and how to teach, evaluate performance, provide feedback, and coordinate interaction with other system models.It decides whether a concept, lesson, or test should come next.Source: §6.3 · p. 15

6.3.1 Decision Making in Cognitive Tutor and Constraint Based Systems.

• Decision strategies range from MTT and CBM feedback to SQL-Tutor recommendations from error history and DEPTHS curriculum replanning from student performance and expert-model sequence.MTT feedback includes correctness flags, buggy messages, and hint chains; DEPTHS rebuilds a concept plan when performance changes significantly.Source: §6.3.1 · pp. 15–16
• Critics argue MTTs need a strategic tutor and should use dialogue to help students construct knowledge rather than simply tell it to them.The criticism contrasts knowledge construction through dialogue with being directed toward knowledge through problem-solving hints.Source: §6.3.1 · p. 16

6.3.2 Tutorial Dialog in Natural Language.

• AutoTutor combines human- and research-derived tutoring strategies with discourse and LSA, while Atlas plans coherent dialogue and DeepTutor models learning progressions.AutoTutor uses strategies including Socratic tutoring and scaffolding-fading; Atlas uses APE to generate dialogue and CARMEL to analyze answers.Source: §6.3.2 · pp. 16–17
• CIRCSIM-Tutor reported significant pre-test to post-test improvement, and its input understander recognized and responded to over 95% of student inputs.Its natural-language tutoring tactics mimic expert human tutors.Source: §6.3.2 · p. 17
• Dialog-based ITSs can be evaluated by how they classify responses, comprehend contributions, model knowledge, and generate conversationally smooth tutorial dialogue.D’Mello and Graesser used AutoTutor as a case study for these evaluation dimensions.Source: §6.3.2 · p. 17

6.3.3 Spoken Dialogue.

• Spoken dialogue is explored partly because human tutors predominantly use speech; speech may support self-explanation and provide cues for a more accurate student model.Hedges, pauses, and disfluencies can provide information about student understanding.Source: §6.3.3 · pp. 17–18
• ITSPOKE was reported more effective than typed dialogue, but no evidence showed increased student learning; SCoT research notes that small recognition errors can reduce effectiveness.Repeated critical feedback was also reported to have negative effects.Source: §6.3.3 · p. 18

7.1 Affective Tutoring System

• Affective tutors use cameras, microphones, specialized mice, and neuro-headsets to detect signals and infer learners’ emotional states in real time.Their objective is to shift negative states, such as confusion, toward positive states, such as commitment.Source: §7.1 · p. 18
• Body language may best detect boredom and flow, while faces convey confusion and delight [123]; empathetic companions improved reported math understanding and interest [126] and reduced interview stress.Tutors may respond to boredom with engaging tasks and to frustration with encouragement or corrective information.Source: §7.1 · p. 19

7.2 Cultural Awareness in Education

• Across three countries, help-seeking was not substantially transferable [130]; Mohammed and Mohan’s system lets learners adjust cultural preferences in problem descriptions, feedback, images, and hints.Source: §7.2 · p. 19

7.3 Game-based Tutoring Systems

• Game-like math tutoring increased interaction likelihood but needs more tutorial features for higher learning; Writing Pal supports writing phases, and Crystal Island users improved post-test scores.For Crystal Island, the pre- to post-test increase in correctly answered questions was statistically significant.Source: §7.3 · p. 20

7.4 Adaptive Intelligent Web Based Educational System (AIWBES)

• Adaptive web tutors personalize goals, preferences, and knowledge; ELM-ART flags prerequisite readiness and ranks relevant links, while ITS-hypermedia combinations can exceed static hypermedia functionality.ELM-ART can also alert learners to prerequisites and suggest related textbook or manual pages.Source: §7.4 · pp. 20–21

7.5 Collaborative Learning

• CycleTalk and WrenchTalk support teams solving tasks with conversational strategies; students working in pairs learned better than students working individually [152].Source: §7.5 · p. 21

7.6 Data Mining in ITSs

• Data mining targets system gaming, affect classification [159], and logged-interaction discovery of problem spaces through clustering and association rules in RomanTutor [161][162][163].Source: §7.6 · pp. 21–22

7.7 Authoring Tools

• Recent paradigm-specific tools aim to reduce programming barriers and development costs: CTAT, ASPIRE, ASAT, and GIFT support model/example tracing, constraint-based tutoring, AutoTutor conversations, and broader services [174][175][176].Source: §7.7 · pp. 22–23

• Because people with special needs may learn more slowly, the authors call for ITS research on targeted skills such as reading and writing, using proven teaching strategies.They suggest such systems may assist learners with conditions including Down Syndrome, traumatic brain injury, dementia, dyslexia, attention deficit disorder, and dyscalculia.Source: §8 · p. 23
• Long-term cohort data could follow students from early ITS use through college, recording variables such as major and GPA, but collecting it may be prohibitively expensive.Source: §8 · p. 23
• The authors report no such datasets yet and propose tracking early math-ITS learners through graduation to examine later outcomes and potentially inform major choices.They suggest the approach might extend to other disciplines and provide time to build skills relevant to students’ eventual majors.Source: §8 · p. 23

• The gap between human tutors and software tutors is narrowing, but remains far from closed.The authors note that knowledge, teaching-style, and student-knowledge models have different strengths and shortcomings, with hybrid models developed to enhance traditional approaches.Source: §9 · p. 23
• ITSs show promise but retain limitations; none has human tutors’ cognitive awareness, though availability, readiness, and consistency may make them competitive where cost, time, and scale matter.Source: §9 · p. 24