Reinforcement Learning in Education: A Systematic Literature Review

Synthesis: Reinforcement Learning in Education: A Systematic Literature Review — A PRISMA-standard systematic literature review of Reinforcement Learning (RL) applications in education, synthesizing 89 manuscripts published 2000–2024 across IEEE Xplore, Google Scholar, and ACM. The authors map the educational contexts, evaluation strategies, RL algorithms, and adaptation mechanisms used in the field, finding promising but methodologically uneven evidence that RL can support Adaptive Learning and Personalized Learning in contexts such as Intelligent Tutoring and Math Education. They report that while over half of studies conducted no statistical testing, those that did frequently showed RL policies outperforming baselines, and they distill a set of best practices for implementing RL in educational practice.

Key Findings

• Scope and method: Following the PRISMA standard, the review analyzed 89 peer-reviewed manuscripts published between 2000 and 2024, drawn from IEEE Xplore, Google Scholar, and ACM, and organized findings across six outcome domains (educational context, evaluation strategy, considered concepts and results, type of algorithm, and type of adaptation).
• Growth and focus areas: Research on Reinforcement Learning in education has grown sharply, with 72 of 89 papers published from 2016 to 2024. Applications concentrate on STEM subjects (39 papers), especially Math Education (n = 27) and Language Learning (n = 11), with the largest target group being college and Higher Education learners (n = 30).
• Algorithms: A majority of papers used model-free RL (n = 72), with value-based methods dominant and Q-learning the most frequently reported algorithm (n = 28). Despite a rise in Deep RL (DRL), classical RL approaches were more consistently effective — 61% of classical RL papers showing significant superiority versus only 36% of DRL approaches.
• Adaptation mechanisms: Adaptation split into content-related (instructional sequencing / content scheduling; n = 53) and guidance-related (hints, Feedback, activity selection; n = 36). RL demonstrated statistically significant superiority over baselines more often for guidance-related tasks than for content scheduling.
• Evaluation and results: Roughly half of papers evaluated "live" with real learners (n = 41), with the rest using real interaction datasets or Simulation of user behavior. Among the 35 papers conducting statistical tests, 18 found at least one RL policy significantly outperforming all baselines; RL appeared comparatively most effective for affective variables (63% significant), followed by task completion, learning/performance, and time on task.
• Effect sizes: Among 15 papers suitable for effect-size analysis, most significant findings reported intermediate to large effect sizes (Cohen's d), suggesting practical significance — though computed from live evaluations only.
• Reward design: Learning gain, especially normalized learning gain, was the most effective reward source (used in over half of papers reporting significant results), sometimes combined with penalties or engagement signals.
• Methodological concerns: Over half of all included papers (n = 54) did not statistically analyze their results, and few used non-adaptive control groups or longitudinal designs, so the growing volume of research is not yet matched by methodological rigor. Publication-bias analyses (funnel plot, Egger's test, PET-PEESE) found no statistically significant evidence of bias but had limited power (n = 6).
• Best practices: The review recommends expanding beyond STEM and younger learners, considering application context, leveraging model-free RL for Adaptive Learning, preferring classical RL over DRL where appropriate, incorporating learning gain into reward functions, prioritizing real-world testing against reasonable baselines (including expert-designed and non-adaptive controls), rigorous statistical evaluation, and longitudinal studies of sustained learning.

What this means for practice

• Designers. Put learning gain — normalized learning gain in particular — in the reward function; it was the most effective reward source across the papers that reported significant results, with penalties or engagement signals added alongside rather than in its place.
• Designers. Prefer classical model-free RL (Q-learning was the most frequently used algorithm, n = 28) before reaching for deep RL: significantly more classical RL papers (61%) than DRL papers (36%) achieved superiority over their baselines.
• Designers. Target guidance before content scheduling — hints, Feedback, and activity selection — since RL showed statistically significant superiority more often for guidance-related adaptation than for instructional sequencing.
• Researchers. Design the comparison properly: RL policies outperformed only about half of expert-designed and non-adaptive baselines, so report non-adaptive and expert baselines alongside algorithmic ones instead of comparing only against weaker or random policies.
• Researchers. Evaluate with real learners and over time, because roughly half of the reviewed papers evaluated live (n = 41) while others used interaction datasets or simulated students, and the field lacks longitudinal evidence that short-term interaction produces sustained learning.

Limitations

• No meta-analysis could be performed: over half of the 89 papers (n = 54) reported no statistical test, most used multiple baselines and dependent variables, and within-study covariances were unavailable, so effect sizes were computed for only 15 papers, all from live evaluations.
• Publication-bias checks had almost no statistical power — only 6 of 89 studies were eligible — and heterogeneity was extreme (Q(5) = 1407.74, p < .001, I² > 75%), so the authors state that publication bias cannot be definitively ruled out and their PET-PEESE result (t(4) = 1.34, p = .251) should be read with caution.
• One author selected and screened the initial 184 records while only the final 89 were double-screened, and the authors acknowledge that personal bias and subjective judgment may have shaped selection and synthesis; restricting the search to peer-reviewed literature may have excluded relevant gray literature and studies using different terminology.
• The corpus is unevenly distributed: 72 of 89 papers appeared between 2016 and 2024, applications cluster in STEM (39 papers) and college learners (n = 30), nearly half of the papers target no specific learner group, and 30% do not specify a learning environment topic.

Citation

Riedmann, A., Schaper, P., & Lugrin, B. (2025). Reinforcement Learning in Education: A Systematic Literature Review. International Journal of Artificial Intelligence in Education, 35(4), 2669–2723.