Bloom’s 2 Sigma Problem: why one-on-one tutoring works, and whether AI can deliver it

Short answer: in 1984 the educational psychologist Benjamin Bloom reported that students who received one-on-one tutoring scored about two standard deviations (2 sigma) above students in a conventional classroom, better than 98% of the classroom group. The catch is that one-on-one tutoring is too expensive to give everyone. The “2 Sigma Problem” asks whether we can find teaching methods that cost about as much as a classroom but work about as well as a tutor.

What the original studies did

Bloom’s paper in Educational Researcher drew on dissertation studies by two of his doctoral students, Anania (1982, 1983) and Burke (1984):

• Students and content: grades four, five and eight, learning probability and cartography. Each sub-study ran for 11 class periods over three weeks.
• Three conditions, with students randomly assigned:
  1. Conventional: about 30 students per teacher, with periodic tests for grading.
  2. Mastery learning: the same class size and instruction, plus formative tests followed by feedback, corrective work on whatever was not yet mastered, and a parallel test to confirm it.
  3. Tutoring: a good tutor for each student (or two or three), with the same formative tests, feedback and corrections.

Results:

• The average tutored student scored about 2 standard deviations above the average conventional student, above 98% of the control group.
• The average mastery-learning student scored about 1 standard deviation higher, above 84% of the control group.
• About 90% of tutored students and 70% of mastery-learning students reached the level that only the top 20% of the conventional class reached.
• Time on task was 65% in the conventional class, 75% under mastery learning and over 90% under tutoring.
• The correlation between aptitude and final achievement fell from +0.60 in the conventional class to +0.35 under mastery learning and +0.25 under tutoring. In other words, with good teaching, starting ability mattered less.

Bloom’s conclusion was that most students can reach a high level of learning; what they lack are the right conditions.

Does the 2 sigma figure hold up?

Honestly: later research has generally not reproduced an effect that large.

• VanLehn (2011) reviewed experiments comparing human tutoring, intelligent tutoring systems and no tutoring. Human tutoring had an effect size of about 0.79 standard deviations over no tutoring; step-based intelligent tutoring systems reached about 0.76, nearly as good.
• Nickow, Oreopoulos and Quan (2020) pooled 96 randomized evaluations of PreK–12 tutoring programs and found an average effect of about 0.37 standard deviations (0.29 in the revised version of the paper).
• The original studies were small, lasted three weeks, and the tutoring group also had mastery learning’s tests and corrections, so not all of the effect can be credited to one-on-one attention.

A fairer summary: one-on-one tutoring is among the most effective forms of teaching we know, usually somewhere between a few tenths of a standard deviation and close to one. Two sigma is a result under very good conditions, not a guarantee.

The ingredients that actually matter

In the same paper, Bloom ranked “alterable variables” that affect achievement by effect size (adapted from Walberg, 1984). The ones closest to tutoring:

1. Feedback and correction (mastery learning), about 1.0. Test after each unit, fix what was missed, confirm it, then move on.
2. Cues and explanations, about 1.0. Make clear what is being learned, why and how.
3. Prerequisites in place, about 0.6. Fill in what the next unit depends on before starting it.
4. Individual attention. A tutor sees where you are stuck and changes the explanation, the examples or the pace. That is why over 90% of tutoring time was spent actually learning.

Bloom’s own proposal was to combine two or three of these variables in group instruction to approach the 2 sigma effect.

Can AI solve the 2 Sigma Problem?

AI makes “one tutor per student” cheap, but having an AI is not the same as having a tutor:

• Chat without teaching design can hurt learning. Bastani et al. (2025, PNAS) found that high school students who practised with a standard chat interface scored 17% lower on the exam, once the AI was removed, than students who never had it.
• AI tutors built on teaching principles can work very well. Kestin et al. (2025, Scientific Reports) found in a randomized crossover study in a Harvard physics course that median learning gains with an AI tutor were more than double those from in-class active learning.

What matters is building in the ingredients above: a clear mastery goal, regular checks, adjusting to the learner’s answers, and bringing back what was learned before.

How Bloom builds this in

Bloom is named after this paper. Here is what it does:

• A mastery goal: the syllabus is a list of things to master, at three depths: Simple (about 8–10 key points), Standard (about 10–12) or In-depth (about 12–15).
• Formative checks: every lesson ends with 2–3 reflection questions, and the next lesson opens by reviewing your answers.
• Adjusting to you: the passages you highlight and ask about, and the feedback you write, decide how the next lesson is written.
• Bringing it back: each lesson starts by reviewing the last one, and earlier ideas return in new examples.
• A finish line: when the syllabus is covered, you get an assessment and a course summary.

Bloom can’t promise two sigma, and no tool can. What it does is put the evidence-backed parts of tutoring within everyone’s reach. Start a course.

References

• 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.
• VanLehn, K. (2011). The relative effectiveness of human tutoring, intelligent tutoring systems, and other tutoring systems. Educational Psychologist, 46(4), 197–221.
• Nickow, A., Oreopoulos, P., & Quan, V. (2020). The impressive effects of tutoring on PreK-12 learning: A systematic review and meta-analysis of the experimental evidence. NBER Working Paper 27476.
• Bastani, H., et al. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. PNAS, 122(26).
• Kestin, G., et al. (2025). AI tutoring outperforms in-class active learning. Scientific Reports, 15, 17458.