Mark ở vũ trụ khác · Another Universe Papers That Changed How We Learn #8 · Tutoring · A classic paper

The 2 sigma problem

Benjamin S. Bloom · Educational Researcher · 1984

The paper"The 2 Sigma Problem: The Search for Methods of Group Instruction as Effective as One-to-One Tutoring"
Typea synthesis built on two University of Chicago dissertations
Main resultthe average tutored student beat 98% of students in a regular class
Mark's labelCaution: right direction, but no one has replicated 2 sigma

Source: Bloom (1984), Educational Researcher 13(6), 4–16; von Hippel (2024), Education Next. The original was checked through excerpts and summaries in search results; the full text could not be opened.

One-minute recap

Essentials

Three ways to teachA regular class of about 30, a "mastery learning" class with quizzes and corrections, and one-to-one tutoring.
2 sigmaThe average tutored student beat 98% of the regular class, according to Bloom.
A problem to solveTutoring every child costs too much. Which whole-class methods come close?
Never replicatedLater syntheses found roughly 0.3 to 0.8, not 2.
The part you can use nowThe "quick quiz, fix mistakes, quiz again" loop needs no tutor.

Source: Bloom (1984); VanLehn (2011), Educational Psychologist; Nickow, Oreopoulos and Quan (2020), NBER w27476; von Hippel (2024), Education Next.

Part 1

The paper's question

If a private tutor can lift an ordinary student to the level of the top of the class, is there a way to teach a whole class that gets close, at a cost society can bear?

Source: the paper's title and question, Bloom (1984), Educational Researcher 13(6). Mark's paraphrase, not a direct quote.

Step 1Compare 3 methodsSame content, same time: regular class, mastery learning, one-to-one tutoring.
→
Step 2Measure the gapIn "sigma", the standard deviation of the regular class's scores.
→
Step 3Pose the problemFind group methods that reach close to 2 sigma by combining several small changes.
Part 2

The discovery story

Not one big experiment. A senior professor wrote up results from two of his students' dissertations and turned them into a question for the whole field.

1956 · A taxonomy of objectivesBenjamin S. Bloom (1913–1999), University of Chicago, is the name behind "Bloom's taxonomy" of learning objectives, from remembering to evaluating.
1968 · "Learning for Mastery"Split a course into units, quiz after each one, and let students who have not yet mastered it correct and relearn before moving on.
Early 1980s · Two dissertationsTwo of Bloom's doctoral students, Joanne Anania and Joseph Arthur Burke, compared three ways of teaching the same content.
1984 · The paper appearsBloom summarizes the two dissertations in Educational Researcher and calls it "the 2 sigma problem".
2011 to 2026 · Measured againVanLehn, then meta-analyses of randomized trials, measure tutoring again. The numbers are much smaller.

Source: education.stateuniversity.com, "B. S. Bloom (1913–1999)" (secondary source, biography, 1956 and 1968); Bloom (1984) (the two students, three conditions); VanLehn (2011); Nickow et al. (2020); Kraft, Schueler and Falken (2026).

The three conditions in the studies Bloom used: a regular class of about 30; a mastery learning class with quizzes and corrections; one-to-one tutoring with the same quiz and correction loop. regular classabout 30 studentsteach, then testfor a grade mastery learningsame class, plusquick quizzes, fixes,quiz again one-to-oneone tutor, one studentsame quiz andcorrection loop same content, same time; only the teaching differs
Three conditions. The two dissertations taught grade 4, 5 and 8 students subjects such as probability and cartography, about 3 weeks each.Source: Bloom (1984) for the three conditions and class size of about 30; grades, subjects and length per von Hippel (2024), Education Next (secondary source; he reread the dissertations).
Part 3

The paper in 3 sentences

The average tutored student scored above 98% of the regular class, about 2 standard deviations higher ("2 sigma"); about 90% of tutored students reached a level only the top 20% of the regular class reached.

The "mastery learning" class, which only added quizzes and corrections, beat about 84% of the regular class, about 1 sigma.

Tutoring every student costs too much, so the problem is to find group methods that come close, for example by combining two or three changes schools can make.

Three score distributions. Regular class centered at 0. Mastery learning shifted 1 sigma right, average above 84 percent of the regular class. Tutoring shifted 2 sigma right, average above 98 percent of the regular class. 0+1σ+2σ final score, in standard deviations of the regular class regular class mastery learningabove 84% of regular tutoringabove 98% of regular Bloom reported: the groups sit 1 and 2 sigma apart
Mark redrew this from the numbers Bloom reported, not from raw data. The graph in the original was itself an illustration, not plotted from individual scores (per von Hippel).Source: Bloom (1984), Educational Researcher 13(6) (98%, 84%, 90% and 20%), via summaries; von Hippel (2024) (secondary source, on the graph).
What do "sigma" and "above 98%" mean?

Sigma (σ) is the standard deviation: how widely scores in the regular class spread. If scores follow a bell curve, a student at the average plus 1 sigma beats about 84% of the class, plus 2 sigma beats about 98%.

So "2 sigma" does not mean double the score. It means an average student, once tutored, jumps to the top of the class.

Source: standard normal-distribution convention; 84% and 98% are how Bloom (1984) himself expressed it.

Part 4

How this paper has been received

Each row has a source and a lookup date.

Where and whenEducational Researcher, volume 13 issue 6, 1984, pp. 4–16, the journal of the American Educational Research Association (AERA). A related version appeared the same year in Educational Leadership.Source: article page (DOI 10.3102/0013189X013006004); ERIC EJ299535.
CitationsGoogle Scholar: about 5,606. Semantic Scholar: 2,869. The databases count differently.Source: the paper's Google Scholar and Semantic Scholar pages, read through search results on 15 October 2026 (counter update date unknown; the true figure may be higher).
Replicated?Not at 2 sigma. VanLehn's review (2011): human tutoring about 0.79 standard deviations over no tutoring. Nickow, Oreopoulos and Quan's meta-analysis of randomized trials (2020): 0.37; 2024 update: 0.288.Source: VanLehn (2011), Educational Psychologist (0.79 via summaries); Nickow et al. (2020), NBER w27476; 2024 version "The Promise of Tutoring for PreK–12 Learning".
Main criticismPaul von Hippel (2024): the original studies were small and short (about 3 weeks), with tests closely matched to what was taught; the graph is an illustration; about half the gain in Anania's study came from extra quizzes and feedback, not the tutor alone. Kraft, Schueler and Falken (2026): tutoring at scale has much smaller effects.Source: von Hippel, "Two-Sigma Tutoring: Separating Science Fiction from Science Fact", Education Next, March 2024 (secondary source, commentary); Kraft et al. (2026), Review of Educational Research.
Awards, textbooksNo reliable source found for awards to this paper, so none is claimed.
VerdictCAUTION Right direction: tutoring and the "quiz, fix" loop really work. The 2 sigma figure comes from a few small, short studies and has never been replicated.The label is Mark's assessment.
Part 5

Fact check: what holds and what is myth

13 things often said about the paper, checked against the paper and later research. Click each row to see why.

SOLIDBloom reported tutored students beat 98% of the regular class

The paper does say so, about 2 sigma; about 90% of tutored students reached the level of the top 20% of the regular class (Bloom 1984). "Solid" means the paper says it, not that the number has been confirmed.

SOLIDMastery learning reached about 1 sigma in the same studies

Bloom reported the class with quizzes and corrections beat about 84% of the regular class (Bloom 1984).

SOLIDThe data came from two doctoral dissertations

Joanne Anania and Joseph Arthur Burke, University of Chicago, students of Bloom (Bloom 1984).

SOLIDThe paper poses a "problem", not a solution

The title: find group methods as effective as one-to-one tutoring. Bloom suggested combining two or three changes that can be made to the learner, the materials, the home and peer group, and the teacher (Bloom 1984, via summaries).

SOLIDTutoring really does improve learning

Meta-analysis of randomized trials: 0.37 standard deviations (Nickow et al. 2020), 0.288 in the 2024 version. Earlier, Cohen, Kulik and Kulik (1982), 65 evaluations of tutoring programs: about 0.40.

CAUTION"Most of the effect is the tutor"

von Hippel (2024) reread Anania's dissertation: about half the gain came from quizzes and extra feedback, which the mastery class also had (secondary source).

CAUTION"It holds for every subject and age"

Original studies: grades 4, 5 and 8, subjects such as probability and cartography, about 3 weeks, tests matched to what was taught (von Hippel 2024, secondary source). Too narrow to generalize.

CAUTION"The three-curve graph is real data"

According to von Hippel (2024), it is a hand-drawn illustration, not plotted from real scores (secondary source).

CAUTION"Tutoring software is nearly as good as a human tutor"

VanLehn (2011): human tutors about 0.79, step-based intelligent tutoring systems about 0.76 (via summaries). VanLehn himself summarized both as about 0.75. But direct comparisons are still few.

CAUTION"AI will solve the 2 sigma problem"

No evidence yet at scale. Papers That Changed How We Learn #10 looks closely at new trials of AI and learning.

FALSE"The 2 sigma effect has been replicated many times"

Later syntheses give roughly 0.3 to 0.8 (VanLehn 2011; Nickow et al. 2020). von Hippel (2024): no trial has replicated 2 sigma.

FALSE"Tutoring at scale works just as well"

Kraft, Schueler and Falken (2026): for large programs measured with independent standardized tests, pooled effects are only about a third to a half as large, falling sharply as programs grow.

FALSE"Bloom said every student needs a tutor"

The opposite: because tutoring every child costs too much, Bloom looked for whole-class methods that come close (Bloom 1984).

Source: as noted in each row; full list in Part 10.

Part 6

What the latest research says

Tutoring works. The question now: how much, and how much is left when thousands of students are served?

Effects in standard deviations. Bloom 1984: 2. VanLehn 2011, human tutors: 0.79. Cohen et al. 1982: 0.40. Nickow et al. 2020: 0.37. 2024 update: 0.29. Programs with 1,000 or more students, Kraft et al. 2026: about 0.16. 012 Bloom 1984, tutoring2 VanLehn 2011, human tutors0.79 Cohen et al. 19820.40 Nickow et al. 20200.37 2024 update0.29 programs ≥1,000 students, 20260.16 average effect (standard deviations), each row from a different synthesis
The more studies and the bigger the scale, the smaller the number. Rows use different comparison groups, so only rough sizes compare. 0.29 is 0.288 rounded.Source: Bloom (1984); VanLehn (2011) (0.79 via summaries); Cohen, Kulik and Kulik (1982), American Educational Research Journal (65 evaluations); Nickow, Oreopoulos and Quan (2020), NBER w27476, and the 2024 version; Kraft, Schueler and Falken (2026), Review of Educational Research (0.16 via a search-result excerpt).

2020 and 2024 · Randomized trials

Nickow, Oreopoulos and Quan pooled tutoring trials from preschool to grade 12: 0.37, and 0.288 in the 2024 version.

2024 · Rereading the dissertations

von Hippel: small, short studies, narrow tests; about half the gain came from quizzes and feedback. "2 sigma" is best read as an ideal ceiling, not an expectation.

2026 · At scale

Kraft, Schueler and Falken, over 250 randomized trials: the bigger the program, the smaller the effect, down to a third to a half. Still meaningful, but far from 2 sigma.

Source: Nickow, Oreopoulos and Quan (2020), NBER w27476, and the 2024 version (EdWorkingPapers); von Hippel (2024), Education Next (secondary source); Kraft, Schueler and Falken (2026), Review of Educational Research, DOI 10.3102/00346543261446660 (trial counts differ between versions, 263 to 282, hence "over 250").

Part 7

Assessing the paper

Scores are Mark's assessment based on the sources in Part 10, not data from a study.

Headline number holds
2/5
Influence
5/5
Easy to apply now
4/5
Strength of original data
2/5
Strengths
  • Asked a clear, measurable question the field has chased for 40 years
  • Showed the "quiz, fix" loop alone was worth about 1 sigma
  • Right direction: tutoring really works
Weaknesses
  • Built on a few small, short studies with narrow tests
  • The graph is an illustration, easily mistaken for data
  • 2 sigma has been quoted as fact for 40 years

A question better than its answer. Read it for the power of fast feedback, not to believe a tutor makes anyone top of the class.

Part 8

Lessons: be your own "tutor" with a quiz-and-fix loop

The most solid part of the paper needs no hired help: a fast feedback loop.

The four-step mastery loop: learn a small unit; take a short quiz; fix exactly what was wrong; quiz again before moving on. 1 · learna small unit 2 · quiz5 questions, closed book 3 · fixrelearn just that part 4 · quiz againpass, then move on →→→ not yet? back to step 3; passed? on to the next unit the "mastery learning" part Bloom measured at about 1 sigma
You do not need a tutor for this loop. A study partner, end-of-chapter questions, or someone to quiz you is enough to start.Source: the mastery loop per Bloom (1984) and "Learning for Mastery" (1968); the concrete steps are Mark's suggestions.
Split it up, quiz after each partAfter each section, answer 5 questions without the book. Under 4 out of 5? Do not move on yet.
Study in pairs, quiz each otherEach writes 5 questions for the other. The asker points out mistakes; the answerer fixes them. The cheapest form of tutoring.Source: Cohen, Kulik and Kulik (1982) include many peer-tutoring programs, average effect about 0.40.
If you hire a tutorAsk them to question and correct more than re-explain. A good session is one where you talk and work more than the tutor.
Realistic expectationsGood tutoring typically helps by about 0.3 to 0.8 standard deviations. That is already a lot. Do not wait for a 2 sigma miracle.Source: Nickow et al. (2020); VanLehn (2011).
Check yourself: do you have a feedback loop? 5 questions

1. The last time I got something wrong, how soon did I find out: minutes, days, or at the exam?

2. For each chapter, do I have a set of questions to answer without the book?

3. When I get something wrong, do I relearn just that part, or reread the whole chapter?

4. Is there someone (a friend, teacher, family member) who could quiz me 10 minutes a week?

5. After 2 weeks, redo the questions I missed: how many are still wrong? Write it down to compare.

Part 9

Further reading

von Hippel (2024), "Two-Sigma Tutoring: Separating Science Fiction from Science Fact"Education Next. A readable reread of the two original dissertations.
VanLehn (2011), "The relative effectiveness of human tutoring, intelligent tutoring systems, and other tutoring systems"Educational Psychologist. Human tutors versus tutoring software.
Nickow, Oreopoulos and Quan (2020), "The Impressive Effects of Tutoring on PreK-12 Learning"NBER Working Paper 27476. A meta-analysis of randomized trials.
Papers That Changed How We Learn #7: Grading 10 study methodsPractice testing, a "high utility" method, is also the heart of the mastery loop.
Papers That Changed How We Learn #9: AI that predicts when you will forgetSettles and Meeder (2016): a computer schedules reviews for each learner.
Part 10

Sources

The main sources are the paper itself and peer-reviewed research. The working environment blocks direct downloads from journal sites, so content was checked through excerpts of those same pages and summaries in search results (looked up 15 October 2026), not the full text.

  1. 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.org/10.3102/0013189X013006004
  2. Bloom, B. S. (1968). Learning for mastery. Evaluation Comment, UCLA.
  3. von Hippel, P. T. (2024). Two-sigma tutoring: Separating science fiction from science fact. Education Next, March 2024. educationnext.org (commentary, used as a secondary source for dissertation details)
  4. Cohen, P. A., Kulik, J. A., and Kulik, C.-L. C. (1982). Educational outcomes of tutoring: A meta-analysis of findings. American Educational Research Journal.
  5. VanLehn, K. (2011). The relative effectiveness of human tutoring, intelligent tutoring systems, and other tutoring systems. Educational Psychologist.
  6. Nickow, A., Oreopoulos, P., and 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. nber.org · 2024 version: "The Promise of Tutoring for PreK–12 Learning".
  7. Kraft, M. A., Schueler, B. E., and Falken, G. (2026). What impacts should we expect from tutoring at scale? Exploring meta-analytic generalizability. Review of Educational Research. DOI 10.3102/00346543261446660. Working paper: EdWorkingPaper 24-1031.
  8. Bloom biography: education.stateuniversity.com, "B. S. Bloom (1913–1999)" (secondary source).
  9. Citation counts: the paper's Google Scholar and Semantic Scholar pages, via search results, looked up 15 October 2026.