Guided Learning and Tutoring: What Randomized Trials Measured

A 2024 meta-analysis of 89 trials found tutoring raised learning by 0.288 SD; a 2026 analysis of 263 trials found 0.16 to 0.22 for large-scale programs.

Last verified: October 20266 min read

Guided one-on-one or small-group instruction, usually called tutoring, has the strongest experimental record of any approach to personalized learning. A 2024 meta-analysis of 89 randomized controlled trials from pre-kindergarten through grade 12 found an average gain of 0.288 standard deviations. A 2026 re-analysis of 263 trials reports 0.40 across its full sample but only 0.16 to 0.22 when the studies resemble large-scale programs judged on independent standardized tests. The headline number therefore depends on which version is cited and on how big the program is.

Key takeaways

  • The 2024 meta-analysis by Nickow, Oreopoulos, and Quan pooled 89 randomized trials and reported 0.288 standard deviations overall, 0.290 for literacy and 0.268 for math.
  • The 2020 working paper reported 0.37 standard deviations, a higher figure from an earlier set of studies and methods.
  • A June 2026 meta-analysis by Kraft, Schueler, and Falken covers 263 randomized trials and finds effects shrink as programs scale up.
  • Effects were larger with teacher or paraprofessional tutors, in earlier grades, with at least three sessions a week, and during school hours.
  • Benjamin Bloom's often-quoted "two sigma" result of 1984 came from small experiments, and evidence on AI tutoring is separate and thinner.

What "guided" instruction means in the research

The research literature rarely uses the phrase "guided education." The closest well-tested form is tutoring, defined in the 2020 working paper as one-on-one or small-group instruction delivered by teachers, paraprofessionals, volunteers, or parents. A whole-class lesson cannot do what a tutor does: an adult watches one learner or a few learners work, corrects errors as they happen, and adjusts the next step. That is the practical meaning of personalization in most trials. Pace and the next task change with the learner's current performance, while the curriculum stays the same.

The authors, Andre Nickow, Philip Oreopoulos, and Vincent Quan, limited the review to randomized field experiments, which means learners were assigned by chance to tutoring or to a comparison group. That design removes most of the guesswork about whether motivated families simply chose tutoring for their children.

Three versions of the evidence

The first version appeared as NBER Working Paper 27476 in July 2020 and reported an overall pooled effect of 0.37 standard deviations. The journal version, "The Promise of Tutoring for PreK-12 Learning," appeared in the American Educational Research Journal (volume 61, issue 1, pages 74 to 107) and was first published online on November 27, 2023, so it is usually cited as 2024. It reports 0.288 (standard error 0.029) across 89 randomized trials. The pages consulted do not say why the figure fell, so the difference cannot be explained here.

The third version is newer. Matthew Kraft, Beth Schueler, and Grace Falken published an analysis of tutoring at scale in Review of Educational Research on June 5, 2026. They expanded the analysis to 263 randomized trials and asked how the estimate changes when the sample is matched to a policy target: large-scale tutoring programs judged on independent, standardized tests. Pooled effects from the studies closest to that target were 0.16 to 0.22 standard deviations, against 0.40 in the full sample. The authors tie the drop to a steep decline in effects as program scale rises, and they report that a bundle of recommended design features partly protects programs from that decline.

SourceEvidence baseReported effectCaution
Nickow, Oreopoulos, and Quan, working paper, July 2020Randomized tutoring trials, preK-120.37 standard deviations overallEarlier estimate, superseded by the journal article
Nickow, Oreopoulos, and Quan, journal article, 202489 randomized trials0.288 overall; 0.290 literacy; 0.268 mathAverage across many program designs
Kraft, Schueler, and Falken, June 2026263 randomized trials0.40 full sample; 0.16 to 0.22 for large-scale, independent-test studiesDifferent sample and methods from the 2024 article
Bloom, Educational Researcher, 1984Small experiments on mastery learning plus tutoringAbout two standard deviations, per secondary accountsNot a large-scale trial; figure known mainly secondhand

Which program features go with larger gains

Four features recur in the 2024 article. Programs staffed by teachers or paraprofessionals had larger average effects than programs using nonprofessional or parent tutors. Effects were larger in earlier grades. Programs that met at least three days a week did better than those that met less often. Tutoring during the school day outperformed tutoring after school, a pattern the 2020 paper also reports. The 2020 paper added that reading tutoring tends to work best in earlier grades and math tutoring in later grades, although overall reading and math effects were similar.

These are averages across trials, and a single program can differ from them. A parent-led program with a well-designed script could outperform a poorly run professional one. The pattern suggests that scheduling and staffing, and not only content, decide results, and the 2026 analysis adds that size matters too.

The two-sigma figure and why it is quoted so often

In 1984, the educational psychologist Benjamin Bloom published "The 2 Sigma Problem: The Search for Methods of Group Instruction as Effective as One-to-One Tutoring" in Educational Researcher (volume 13, issue 6, pages 4 to 16). Secondary accounts summarize it this way: students taught by mastery learning in a group did about one standard deviation better than a conventional class, and students tutored one-on-one with mastery methods did about two standard deviations better. Bloom posed the challenge of finding group methods as effective as tutoring.

The original article could not be read in full for this page, so the figures above rest on secondary descriptions and are not verified against the paper itself. They came from small studies. Every pooled estimate from randomized trials, from 0.40 down to 0.16, is a fraction of two standard deviations, which is a more realistic expectation for a school system that funds tutoring.

Questions to ask about a tutoring or guided-learning program

  • Who delivers the sessions, and what training or supervision do they have?
  • How many sessions a week, and are they inside the school day?
  • How many students does the program serve, and was it tested at a similar size?
  • Was the program tested with a randomized or well-matched comparison group, and in which grades?
  • Is the outcome a standardized test, a program-made test, or a grade?
  • If software or AI supplies the guidance, was that specific tool tested, rather than human tutoring in general?

This article summarizes published research for general information. It does not recommend a program or a provider, and school or district decisions belong with the educators and administrators who know the local students.

FAQ

Is tutoring the same as personalized learning?

Not exactly. Personalized learning is a broad label that can cover software, flexible pacing, or student choice. Tutoring is one specific form with a large randomized evidence base, which is why it is the best-documented example.

Why do some sources say 0.37, others 0.288, and others 0.40?

The 0.37 figure comes from the July 2020 working paper. The 2024 journal version reports 0.288 across 89 trials. The 0.40 figure is the full-sample estimate of the June 2026 analysis of 263 trials, which also reports 0.16 to 0.22 for studies resembling large-scale programs on independent tests.

Does the evidence apply to AI tutors?

Not directly. The trials reviewed here involved human tutors. Studies of AI chatbot tutors are fewer, use different designs, and need their own evaluation. No pooled estimate for AI tutors was checked for this page.

Does more tutoring always mean more learning?

The meta-analysis found larger average effects when sessions happened at least three times a week, but that is an association across programs, not a guarantee for any single student.

Last verified: October 2026

Sources

  • The Impressive Effects of Tutoring on PreK-12 Learning, Nickow A, Oreopoulos P, Quan V, NBER Working Paper 27476, July 2020
  • The Promise of Tutoring for PreK-12 Learning: A Systematic Review and Meta-Analysis of the Experimental Evidence, Nickow A, Oreopoulos P, Quan V, American Educational Research Journal 61(1), 2024
  • Exploring Meta-Analytic Generalizability (subtitle of the article on tutoring impacts at scale), Kraft MA, Schueler BE, Falken GT, Review of Educational Research, June 2026
  • The 2 Sigma Problem: The Search for Methods of Group Instruction as Effective as One-to-One Tutoring, Bloom BS, Educational Researcher 13(6), 1984 (ERIC record EJ303699)

New to artificial intelligence?

Start with the basics, then follow the guides to tools, rules, and effects on society.

Start with AI Basics