Mastery learning
Mastery learning is a way of teaching in which material is split into short units, learners take a short check after each unit, and those who have not yet reached a set standard get corrective help and a second check before moving on. The idea rests on John B. Carroll's 1963 model of school learning, which treats aptitude as the amount of time a learner needs, and was turned into a classroom method by Benjamin S. Bloom in his 1968 paper "Learning for Mastery" [1] [2]. Bloom argued that most students, perhaps more than 90 percent, can master what is taught if instruction and time are matched to their needs [2].
In 1984 Bloom reported that students taught one-to-one by a tutor, with the same mastery checks, scored about two standard deviations (two "sigma") above students in ordinary classes: the average tutored student did better than 98% of the comparison class [3]. He called the search for group methods that could match this result the "2 sigma problem". Later reviews found smaller effects for human tutoring, about 0.8 standard deviations, and similar effects for the best computer tutors [4].
Contents
Key facts
- Carroll's 1963 model treats a learner's aptitude as the time needed to learn a task, so the degree of learning depends on the time actually spent relative to the time needed [1] [2].
- In Bloom's 1984 summary, the average student under mastery learning scored about one standard deviation above the conventional class (above 84% of it), and the average tutored student about two standard deviations above (above 98%) [3].
- About 90% of tutored students and 70% of mastery learning students reached the final test level that only the top 20% of conventional students reached [3].
- A 1990 meta-analysis (a study that pools earlier studies) of 108 controlled evaluations found positive effects of mastery programs on exam performance, stronger for weaker students [5].
- A 1987 review of group-based mastery learning in schools found essentially no effect on standardized tests and moderate effects on tests written by the researchers [6].
- A 2011 review put the effect of human tutoring at 0.79 standard deviations and that of intelligent tutoring systems (computer tutors that respond to each step a learner takes) at 0.76 [4].
Origins of mastery learning
Carroll's model of school learning
Aptitude as learning time
In "A Model of School Learning" (1963), the psychologist John B. Carroll proposed that learners differ less in whether they can learn something than in how long it takes them [1]. Degree of learning is a function of the time a learner actually spends on a task divided by the time the learner needs. Time needed depends on aptitude, the quality of instruction and the ability to understand instruction; time spent depends on the time allowed and on perseverance, the time the learner is willing to spend.
Bloom's five variables
Bloom built directly on these variables. In "Learning for Mastery" he named five: aptitude, measured as the time needed to master a task; quality of instruction; ability to understand instruction; perseverance; and the time allowed for learning, which he treated as the key factor [2].
Bloom's learning for mastery
The 1968 paper
Bloom's 1968 paper appeared in Evaluation Comment, a newsletter on educational evaluation [2]. His starting point was that teaching all students in the same way and giving all of them the same time produces large differences in what they learn. If methods and time were varied to fit each learner, he argued, nearly all could reach a high standard [7].
Three parts of the method
The method he proposed has three main parts [3] [7]:
- Formative assessment: a short test after each learning unit, used for feedback rather than for grades.
- Correctives: extra explanation, different materials or group work aimed at the specific items a learner missed.
- A parallel second test: a new version of the formative test, to check whether the correctives worked before the class moves on.
Mastery standards
Mastery is usually defined as a set score on the unit test; in one study Bloom described, the standard was 80% [3]. Guskey notes that the method was also reported to improve students' confidence, attendance, class participation and attitudes toward learning [7].
Keller's personalized system of instruction
Working independently, the psychologist Fred S. Keller developed a related college course format, described in his 1968 paper "Good-bye, teacher..." [8]. Known as the personalized system of instruction (PSI) or the Keller Plan, it has five features: students move through the material at their own pace; they must show mastery of each unit before going to the next; lectures and demonstrations are used to motivate rather than as the main source of information; written materials carry most of the teaching; and proctors, often more advanced students, give immediate tests, scoring and tutoring [8]. PSI differs from Bloom's version mainly in pacing: Bloom's classes moved together as a group, while PSI students moved individually [8] [9].
Evidence from classroom studies
Major reviews of mastery learning
Reviews of mastery learning in the 1980s and 1990s reached different conclusions depending on which studies they included and how achievement was measured.
| Review | Studies covered | Main finding |
|---|---|---|
| Slavin 1987 | Group-based mastery learning in elementary and secondary schools, at least 4 weeks long | Essentially no effect on standardized tests; moderate positive effects on experimenter-made tests, with little evidence they lasted [6] |
| Guskey and Pigott 1988 | 46 studies of group-based mastery learning | Consistently positive effects on cognitive and affective outcomes, but effect sizes varied widely between studies [9] |
| Kulik, Kulik and Bangert-Drowns 1990 | 108 controlled evaluations | Positive effects on exam performance, stronger for weaker students; more time on tasks; self-paced college programs often had lower completion rates [5] |
| VanLehn 2011 | Experiments comparing human tutors, computer tutors and no tutoring | Human tutoring d = 0.79; step-based intelligent tutoring 0.76 [4] |
| Ma and colleagues 2014 | 107 effect sizes, 14,321 participants | Intelligent tutoring beat large-group teaching (g = 0.42); no significant difference from one-to-one human tutoring [10] |
| Kulik and Fletcher 2016 | 50 controlled evaluations of intelligent tutoring | Median gain 0.66 standard deviations, from the 50th to the 75th percentile [11] |
Coverage versus mastery
Slavin's review placed the debate in terms of a "coverage versus mastery" dilemma: time spent bringing every student to the standard on one unit is time not spent covering new material [6]. The 1990 meta-analysis found related costs, noting that mastery programs may increase the time students spend on tasks and that self-paced programs often reduce completion rates in college classes [5]. Several reviews also found that results depend strongly on the test used: effects are larger on tests written for the course than on standardized tests [6] [11].
The 2 sigma problem
Bloom's 1984 tutoring studies
Bloom's 1984 paper in Educational Researcher drew on dissertations by two of his University of Chicago doctoral students Anania and Burke [3]. Students were randomly assigned to three conditions: a conventional class of about 30 students with periodic tests for grades; a mastery learning class of about 30 with formative tests, correctives and parallel tests; and tutoring, with a good tutor for each student or for two or three students at once, plus the same formative tests and correctives. The studies covered students in grades 4, 5 and 8 and two subjects, probability and cartography, and each sub-study lasted 11 periods over 3 weeks [3].
Reported results
The results Bloom reported were [3]:
- The average tutored student scored about two standard deviations above the average of the conventional class, above 98% of its students.
- The average mastery learning student scored about one standard deviation above, above 84% of the conventional class.
- About 90% of tutored students and 70% of mastery learning students reached the level of the top 20% of the conventional class.
- Time on task in class was about 65% under conventional teaching, 75% under mastery learning and above 90% under tutoring.
- The correlation between prior aptitude and final achievement fell from about +0.60 in conventional classes to +0.35 under mastery learning and +0.25 under tutoring.
The challenge for group teaching
Bloom argued that one-to-one tutoring is too costly to provide at scale, and defined the "2 sigma problem" as finding group methods that could reach similar results [3]. He combined mastery learning with other variables from a table of effect sizes. One example came from a study by Leyton in Algebra 2 and French 2 classes: reviewing the prerequisites from the earlier course at the start of the term, combined with mastery learning, put the experimental group about 1.6 standard deviations above the control group on the final exam [3]. Bloom also noted that the outstanding performers in his talent study had received much of their teaching one-to-one, with feedback and correctives, see Stages of talent development [12].
Criticism of the two sigma figure
The two-sigma figure* rests on small, short studies (three-week units in two subjects), and it was not matched in later reviews. VanLehn wrote that a human tutoring effect of about 2.0 had become a widely held belief, but his review found 0.79 [4]. Slavin noted that Bloom's prediction of two-sigma gains from mastery-based methods was not supported by school-based studies on standardized tests [6].
Tutoring and language courses
Computer and AI tutoring
Intelligent tutoring systems
Intelligent tutoring systems, computer programs that track a learner's work step by step and give targeted hints, have long been studied as a partial answer to the 2 sigma problem. VanLehn found that step-based systems were nearly as effective as human tutors, both well below the two-sigma figure [4]. A meta-analysis by Ma and colleagues found intelligent tutoring better than teacher-led large-group instruction, other computer-based instruction and textbooks, and not significantly different from one-to-one human tutoring [10]. Kulik and Fletcher found a median gain of 0.66 standard deviations across 50 evaluations, but small effects where control groups were unusual or the systems were poorly implemented [11].
Tutors built on language models
Tutors built on large language models are a newer attempt in the same direction, and the evidence is still early. In a 2023 randomized crossover trial with 194 students in an introductory physics course at Harvard, students learned more from an AI tutor designed around the course's teaching practices than from an in-class active learning session, with effect sizes of 0.73 to 1.3 standard deviations, and a median time on the AI lessons of 49 minutes [13]. The authors noted a ceiling effect on the post-test and wrote that they did not expect structured AI tutoring to outperform in-class learning in all contexts [13]. In a World Bank randomized trial in Nigeria, six weeks of after-school tutoring with Microsoft Copilot (based on GPT-4) for English language learning in the first year of senior secondary school raised English scores by 0.23 standard deviations, with larger gains for girls and for students with higher starting scores [14]. Both are single studies, in different subjects and settings, and neither tested a long course.
Application to language learning
Sequenced language courses
Mastery learning fits language courses that are built as sequences, where each unit depends on earlier ones. Bloom's own example of a large combined effect came from a French 2 course [3]. In practice, a formative check might test the forms of Ser and estar after a unit, with corrective practice on the specific errors before the next unit. Corrective feedback on a learner's own speech and writing is discussed under Output and corrective feedback.
Limits of strict mastery criteria
Languages also pose a problem for strict mastery criteria. Fluency builds gradually over many encounters rather than unit by unit, so many courses combine mastery checks with recycling of earlier material through Spaced repetition and Retrieval practice. Choosing the next unit at the edge of what a learner already knows is covered under Knowledge frontier, and the levels of thinking that mastery tests can target are classified in Bloom's taxonomy. Listening and reading material that learners mostly understand while still meeting new language is the subject of Comprehensible input.
See also
Pages that link here
References
- Carroll, J. B. (1963). A Model of School Learning. Teachers College Record, 64(8), 723-733. doi:10.1177/016146816306400801
- Bloom, B. S. (1968). Learning for Mastery. Evaluation Comment, 1(2) (ERIC ED053419). eric.ed.gov/?id=ED053419
- Bloom, B. S. (1984). The 2 Sigma Problem: The Search for Methods of Group Instruction as Effective as One-to-One Tutoring. Educational Researcher, 13(6), 4-16. doi:10.3102/0013189X013006004
- VanLehn, K. (2011). The Relative Effectiveness of Human Tutoring, Intelligent Tutoring Systems, and Other Tutoring Systems. Educational Psychologist, 46(4), 197-221. doi:10.1080/00461520.2011.611369
- Kulik, C.-L. C., Kulik, J. A., Bangert-Drowns, R. L. (1990). Effectiveness of Mastery Learning Programs: A Meta-Analysis. Review of Educational Research, 60(2), 265-299. doi:10.3102/00346543060002265
- Slavin, R. E. (1987). Mastery Learning Reconsidered. Review of Educational Research, 57(2), 175-213. doi:10.3102/00346543057002175
- Guskey, T. R. (2007). Closing Achievement Gaps: Revisiting Benjamin S. Bloom's 'Learning for Mastery'. Journal of Advanced Academics, 19(1), 8-31. doi:10.4219/jaa-2007-704
- Keller, F. S. (1968). Good-bye, teacher.... Journal of Applied Behavior Analysis, 1(1), 79-89. doi:10.1901/jaba.1968.1-79
- Guskey, T. R., Pigott, T. D. (1988). Research on Group-Based Mastery Learning Programs: A Meta-Analysis. The Journal of Educational Research, 81(4), 197-216. doi:10.1080/00220671.1988.10885824
- Ma, W., Adesope, O. O., Nesbit, J. C., Liu, Q. (2014). Intelligent tutoring systems and learning outcomes: A meta-analysis. Journal of Educational Psychology, 106(4), 901-918. doi:10.1037/a0037123
- Kulik, J. A., Fletcher, J. D. (2016). Effectiveness of Intelligent Tutoring Systems. Review of Educational Research, 86(1), 42-78. doi:10.3102/0034654315581420
- Bloom, B. S., Sosniak, L. A. (1981). Talent Development vs. Schooling. Educational Leadership, 39(2), 86-94. files.ascd.org/staticfiles/ascd/pdf/journals/ed_lead/el_198111_bloom.pdf
- Kestin, G., Miller, K., Klales, A., Milbourne, T., Ponti, G. (2025). AI tutoring outperforms in-class active learning: an RCT introducing a novel research-based design in an authentic educational setting. Scientific Reports, 15, 17458. doi:10.1038/s41598-025-97652-6
- De Simone, M., Tiberti, F., Barron Rodriguez, M., Manolio, F., Mosuro, W., Dikoru, E. J. (2025). From Chalkboards to Chatbots: Evaluating the Impact of Generative AI on Learning Outcomes in Nigeria. World Bank Policy Research Working Paper 11125. openknowledge.worldbank.org/bitstreams/cd9cca5b-2ffd-4be1-8313-0f4d2ed6ef4f/download