⚡ AdaptEdu ← Back to app
The evidence — both sides

The research behind how we teach.

AdaptEdu's engine is built on decades of learning-science research. Below is the evidence that supports our approach — and, just as importantly, the evidence that complicates it. We think you should see both.

Why show the counter-evidence? Because a learning company that only cites the studies it likes isn't one you should trust with how people learn. The honest picture is more useful than a flattering one — and it's how we decide what to build. Each card shows the source's own title, a plain-language summary, and a link to the full study so you can check our reading against theirs.
Supports our approach
Complicates the picture
Supports our approach

Evidence the engine is built on

These are the mechanisms our adaptive engine actually uses: testing as a learning event, spacing material over time, mastery before advancing, multiple representations of the same idea, and one-to-one-style adaptation.

Testing isn't just measurement — it's how memory forms

Retrieval practice

"Test-enhanced learning: taking memory tests improves long-term retention."— Roediger & Karpicke, Psychological Science, 2006 · 17(3):249–55 · doi:10.1111/j.1467-9280.2006.01693.x

In plain terms: students who were quizzed on material remembered significantly more weeks later than students who simply reread it the same number of times — even though rereading felt more effective at the time.

Why it matters for usOur engine runs short retention checks after each micro-module instead of saving testing for a final exam. The act of retrieving is the learning, not just a grade.
Read on PubMed →

Spacing material out beats cramming it together

Spaced practice

"Distributed practice in verbal recall tasks: A review and quantitative synthesis."— Cepeda, Pashler, Vul, Wixted & Rohrer, Psychological Bulletin, 2006 · 132(3):354–80

In plain terms: a meta-analysis pooling 839 assessments across 317 experiments found that spreading study out over time produces substantially better long-term retention than the same study packed into one session.

Why it matters for usOur retention checks deliberately reach back to earlier modules rather than only testing what you just saw — spacing the recall instead of bunching it.
Read on PubMed →

One-to-one tutoring with mastery moved the average student to the top 2%

Mastery & tutoring

"The 2 Sigma Problem: The Search for Methods of Group Instruction as Effective as One-to-One Tutoring."— Benjamin S. Bloom, Educational Researcher, 1984 · 13(6):4–16

In plain terms: Bloom reported that students taught one-to-one with mastery learning performed about two standard deviations better than students in conventional classrooms — roughly moving an average student to the top few percent.

Why it matters for usThis is the core case for adaptation: one-to-many instruction leaves most of that gain on the table. Software can't fully replace a human tutor, but it can carry the mastery-and-adapt mechanics to far more people. (See the honest caveat below.)
Read the paper →

People learn more from words and visuals together — through separate channels

Multiple representations

"The Cambridge Handbook of Multimedia Learning."— Richard E. Mayer (ed.), Cambridge University Press · Cognitive Theory of Multimedia Learning

In plain terms: the brain processes words and pictures through separate, limited-capacity channels. Presenting an idea across complementary formats — without overloading either channel — improves understanding and retention for learners broadly.

Why it matters for usThis is why our engine rotates between video, diagram, text, and interactive formats — not because a learner has a fixed "style," but because multiple representations of the same concept help nearly everyone. (Critical distinction — see below.)
Cambridge Handbook →
Complicates the picture

Evidence we take seriously against us

If we ignored this research, we'd be building on sand. These findings shape what we deliberately don't claim — and where we stay humble about results.

"Learning styles" don't hold up — and we don't claim them

Against the field

"Learning Styles: Concepts and Evidence."— Pashler, McDaniel, Rohrer & Bjork, Psychological Science in the Public Interest, 2008 · 9(3):105–19

In plain terms: this landmark review found no adequate evidence for the "meshing hypothesis" — the popular idea that matching instruction to a learner's supposed style (visual, auditory, kinesthetic) improves outcomes. The belief is widespread; the supporting evidence is not.

How it changes what we buildWe explicitly reject learning styles. Our engine does not sort you into a "type" and feed you one format. It adapts on measured retention — what you actually remembered — and uses multiple representations because they help everyone, per Mayer, not because of a style label.
Read on PubMed →

"Personalized learning" at scale produced modest, inconsistent gains

Implementation reality

"Informing Progress: Insights on Personalized Learning Implementation and Effects."— Pane, Steiner, Baird, Hamilton & Pane, RAND Corporation, 2017 (Gates Foundation–funded)

In plain terms: RAND's studies of personalized-learning schools found generally positive but modest and uneven effects, with wide variation between schools and serious implementation challenges. Technology alone did not guarantee results.

How it changes what we buildIt's why we don't promise a silver bullet. The product has to be genuinely usable, the content has to be good, and adoption matters as much as the algorithm. We treat implementation and content quality as first-class problems, not afterthoughts.
Read the RAND report →

The full "2 sigma" effect is hard to reproduce at scale

Caveat on our own claim

A caveat on the Bloom (1984) result we cite above.— Later researchers have rarely reproduced the full two-sigma effect outside Bloom's original conditions.

In plain terms: the two-standard-deviation result is real and influential, but real-world tutoring and adaptive programs typically show meaningful — but smaller — gains than Bloom's headline figure.

How it changes what we buildWe cite 2-sigma as direction, not destination. Our honest internal target is "meaningfully better than one-to-many instruction for the learner who sticks with it," and we plan to measure our own outcomes rather than borrow Bloom's number.
Read Bloom (1984) →

Where this leaves us

The research points in a consistent direction: retrieval, spacing, mastery, and multiple representations work — and adaptation to the individual learner can beat one-size-fits-all instruction. That's the foundation we build on.

But the same literature is clear about what doesn't work: matching content to "learning styles" is not supported, and personalized-learning technology only delivers when the content and the implementation are genuinely good. Those findings are guardrails we designed around, not inconveniences we ignore.

So our claim is deliberately modest and testable: an engine that adapts on measured retention, sequences with spacing and mastery, and teaches concepts through multiple representations should help a committed learner more than traditional one-to-many instruction. We intend to prove that with our own outcome data — and we'll publish what we find, including the parts that don't flatter us.