Bloom’s 2-sigma problem, forty years later.

Some findings organize a field the way a mountain organizes the land around it: everything that comes after is placed relative to the peak. In learning science that peak is a single number from 1984 — endlessly cited, widely misremembered, and more useful when you know exactly what it did and did not measure. This article walks through the original studies. Then the forty-year attempt to replicate their effect at scale. Then the surprisingly hopeful place that attempt has now reached.

In 1984, Educational Researcher published a paper by Benjamin Bloom that gave learning science its most famous number. Two of Bloom’s doctoral students at the University of Chicago had run controlled experiments comparing three ways of teaching the same material: a conventional classroom, a mastery-learning classroom, and one-to-one tutoring. The headline result: the average tutored student performed about two standard deviations — two sigma — above the average of the conventional class. That means better than roughly 98 percent of students taught the ordinary way (Bloom, 1984).

Bloom did not present the number as a victory lap. He presented it as a problem. One-to-one tutoring was, in his words, too costly for most societies to bear on a large scale, so the task he set the field was to find methods of group instruction as effective as a personal tutor (Bloom, 1984). Forty years later, that assignment has structured an entire research programme: mastery-learning systems, intelligent tutoring systems, and now large-language-model tutors are all, in one way or another, attempts to hand in Bloom’s homework.

What Bloom actually measured

The details matter more than the legend. The two-sigma figure comes mainly from dissertation studies by Joanne Anania and Arthur Burke, run with fourth-, fifth- and eighth-grade students learning short, well-structured units such as probability and cartography (Anania, 1983). The three conditions were carefully staged. Conventional classes of about thirty took periodic tests. Mastery-learning classes added formative tests with corrective feedback until each unit was mastered. Tutoring ran in groups of one to three, kept the mastery procedures, and added a tutor.

Two features of the result are routinely forgotten. First, mastery learning alone — no tutor, just formative testing with required correctives — moved students roughly one standard deviation above the conventional class (Bloom, 1984). Half of the fabled gap came from procedures that need no tutor at all. Second, the tutoring condition included the mastery requirement. Tutored students were not merely talked to one-on-one; they were not allowed to move on until they had shown mastery of each unit. The famous 2.0 is the effect of tutoring and mastery together, under near-ideal conditions, measured with tests closely matched to exactly what was taught.

Two sigma was the ceiling, not the norm

The natural next question: does the lab number travel? Here the literature delivers its first correction to the legend.

How typical is that benchmark of tutoring in the wild? Not very. A meta-analysis of school tutoring programmes — real tutors, real classrooms, no laboratory staging — found an average effect on tutored students of about 0.40 standard deviations (Cohen, Kulik & Kulik, 1982): clearly valuable, nowhere near two sigma.

Kurt VanLehn later reviewed experiments that directly compared human tutoring against classroom or no-tutoring baselines. The typical effect of human tutoring: d ≈ 0.79 — large, but far short of Bloom’s 2.0 (VanLehn, 2011). On this reading, Bloom’s studies were an outlier built by stacking every advantage at once: trained tutors, mastery thresholds, short curated units, and outcome tests matched to the instruction. That reframing matters because it moves the finish line. The realistic target for “as effective as a human tutor” is closer to 0.8 standard deviations than to 2.0. And that target, it turns out, is reachable.

What a tutor actually does

The mechanism behind the effect is easy to misread. A tutor’s advantage is not superior explanation — lectures from a tutor are still lectures. It is a loop that group instruction cannot run. The tutor continuously diagnoses the learner’s current state, often from a single wrong step. The tutor guides, with a hint or question tuned to that specific error. And the tutor verifies that the learner can now do the thing before moving on.

Mastery learning captures the verify step alone — which is presumably why it is worth around a full sigma by itself in Bloom’s data (Bloom, 1984). Tutoring adds high-frequency diagnosis and guidance in between.

VanLehn’s synthesis offers the sharpest version of this claim: the interaction granularity hypothesis — in plain terms, how fine-grained the help is. Teaching gets more effective as the feedback loop tightens, from feedback on final answers to feedback on every step of a solution. Then it plateaus; help finer than the step added little in the studies he reviewed (VanLehn, 2011). If that plateau is real, it is extremely good news for software. Step-level interaction is exactly the grain a machine can sustain forever, for every learner at once.

The forgotten half of the result

Before the machines enter the story, it is worth dwelling on the condition everyone skips: mastery learning alone. The procedure was almost bureaucratically simple. Teach a unit to the whole class, then give a short formative test — diagnostic, not graded. Have every student who fell short work through correctives aimed at exactly what they missed; re-test; only then move the class forward. No tutor, no tailoring of the first teaching, no technology. In Bloom’s data this alone lifted the average student roughly a full standard deviation above the conventional class (Bloom, 1984).

That result carries two lessons that survive even the most skeptical reading of the 2.0 headline. First, a large share of what looks like a tutoring effect is really a verification effect. Conventional instruction advances on the calendar, letting gaps compound silently from unit to unit. Simply refusing to advance past unverified material removes the compounding. Second, the intervention operates on the teaching system, not the student — the same learners, teacher, and material produced dramatically different outcomes when the procedure changed.

Bloom drew from this his career-long theme: most students can reach levels of attainment usually seen only in the top few, given conditions that catch and correct errors early (Bloom, 1984). Any software system that gates progression on demonstrated mastery is claiming this half of the legacy — the half that never needed a tutor at all.

How close have machines come?

With the target reset to roughly 0.8, and the mechanism named as a loop rather than a person, the question becomes an engineering one. How much of that loop can software actually run? The answer has been piling up for thirty years, and it is better than the field’s public reputation suggests.

Intelligent tutoring systems (ITS) were built as an explicit run at Bloom’s target. Carnegie Mellon’s cognitive-tutor group even titled a paper “Cognitive computer tutors: solving the two-sigma problem.” From evaluations of its programming and algebra tutors, the group argued that mastery-based cognitive tutors could recover much of the tutoring advantage (Corbett, 2001). These systems model the knowledge a task requires and track each learner’s mastery of each component skill. They give feedback within a solution attempt rather than after it, and they gate progression on demonstrated mastery. The tutoring loop, mechanized.

The meta-analytic verdict is more positive than most people outside the field expect. In VanLehn’s comparison, step-based tutoring systems produced d ≈ 0.76, against the same kinds of baselines on which human tutoring produced 0.79. That difference is too small to mean anything (VanLehn, 2011). Ma and colleagues pooled 107 comparisons. ITS beat conventional large-group instruction by a moderate margin, and showed no significant difference from one-on-one human tutoring (Ma et al., 2014). Kulik and Fletcher, reviewing 50 evaluations of full ITS courses, reported a median effect of 0.66 standard deviations — with the tutoring system beating the comparison condition in the large majority of studies (Kulik & Fletcher, 2016).

Against Bloom’s 2.0, those numbers look like failure. Against the realistic human-tutoring benchmark of roughly 0.8, they read very differently. On the outcomes these studies measured, well-built tutoring software performs in the same band as a human tutor — at a marginal cost per learner approaching zero.

The LLM turn

If step-based systems already match human tutors on measured outcomes, the two-sigma story could have ended a decade ago as a quiet success. It didn’t, because the success never spread — and the reason it never spread is the same kind of economics that motivated Bloom’s challenge in the first place.

The specific economics: authoring. A classic ITS demanded years of cognitive task analysis and rule engineering per course. That is one reason the successes cluster in math, physics and programming (Kulik & Fletcher, 2016). Large language models change that constraint. Tutor-style dialogue no longer requires a hand-built domain model for every topic — though it very much still requires pedagogical guardrails.

The early controlled evidence is promising. A randomized trial ran in a Harvard introductory physics course. Students worked through lessons with a purpose-built AI tutor, engineered around the same principles the ITS tradition converged on: activate prior knowledge, guide rather than answer, check understanding before advancing. They learned substantially more than students in an actively taught classroom, and did so in less time (Kestin et al., 2025). It is one course, one population, and a short duration. But it is exactly the shape of result Bloom’s challenge calls for, and the first trials of this new generation are only now arriving.

What the evidence doesn’t show

This literature is genuinely strong, and it is routinely oversold. Four caveats keep the claims honest:

• Nothing reliably delivers two sigma. No replicated programme — human or machine — has reproduced Bloom’s 2.0 under realistic conditions. Typical effects for tutoring of all kinds cluster between roughly 0.4 and 0.8 (Cohen, Kulik & Kulik, 1982) (VanLehn, 2011). Two sigma is a ceiling observed under ideal conditions, not an achievable KPI.
• Effects shrink on standardized tests. Kulik and Fletcher found markedly larger effects when outcomes were measured with locally developed tests than with standardized ones (Kulik & Fletcher, 2016). Some of that is legitimate alignment, some of it is teaching to the measure — either way, the gains are narrower than a single headline number implies.
• Results vary by population and subject. A meta-analysis restricted to K–12 mathematics found overall ITS effects close to zero, with weaker results for lower-achieving students (Steenbergen-Hu & Cooper, 2013). The averages conceal real heterogeneity.
• Implementation dominates at scale. In the largest randomized effectiveness trial of a mature ITS — Cognitive Tutor Algebra I across dozens of schools — there was no detectable effect in the first year; a positive effect of roughly a fifth of a standard deviation appeared only in the second year of implementation (Pane et al., 2014). The software is a necessary ingredient, not a sufficient one.

Where the evidence stops

1. 1Nothing reliably delivers two sigma
2. 2Effects shrink on standardized tests
3. 3Results vary by population and subject
4. 4Implementation dominates at scale

© 2026 FUTURE PROOF™

None of this undermines the core result; it bounds it. Tutoring-style instruction produces large, replicable gains on outcomes aligned to what was taught. Those gains shrink with distance from the taught material, vary across populations, and depend heavily on how the system is actually used. A buyer who expects two sigma will be disappointed by 0.6. A buyer who knows that 0.6 is what a good human tutor typically delivers, at a fraction of the cost, is reading the same evidence correctly.

Where that leaves the problem

The pedagogical guardrails deserve one more sentence, because they are where the ITS tradition and the LLM generation genuinely meet. A general-purpose model left to its own devices behaves like an eager answer key. It explains fluently and resolves the learner’s difficulty for them — precisely the move the tutoring literature says forfeits the effect. The systems that work constrain the model into the loop the research isolated: withhold the answer, probe the step, verify before advancing. The intelligence is new; the pedagogy it must be bent to is forty years old.

Read at forty years’ distance, Bloom’s paper makes two claims, and history has treated them differently. The empirical claim — two full standard deviations as a typical effect — has not survived; it was the ceiling, not the norm (VanLehn, 2011). The structural claim has aged perfectly: the active ingredients of tutoring are continuous diagnosis, calibrated guidance and verified mastery, and the only reason every learner does not get them is that human loops do not scale. That was never a pedagogy problem. It was an economics problem — and it is the one part of the 1984 paper that technology can straightforwardly attack (Bloom, 1984).