How Do You Learn Best? What the Research Really Says

Most of us learn the way we saw in school: read the lecture handout, highlight the important parts in yellow, read it again, and cram everything into one night before the exam. It feels like work. Unfortunately, precisely this combination is well researched, and it performs poorly.

The good news: the methods that work better are not complicated. They are just less comfortable. Here is what research says about effective learning, with numbers, study designs, and honest notes on where the evidence gets thin.

The short version

If you only read one paragraph, make it this one. In 2013, Dunlosky et al. evaluated ten common learning techniques against the available research. The result:

• High utility: practice testing (that is, retrieving from memory) and distributed practice.
• Moderate utility: elaborative interrogation, self-explanation, interleaved practice (mixing problem types).
• Low utility: summarizing, highlighting and underlining, keyword mnemonics, imagery for text learning, and rereading.

Two caveats apply. First, these are summary assessments, not a shared ranking from one experiment. Second, "low utility" does not mean "useless"; it means limited or inconsistent effects across different learners, materials, and exams. Highlighting and rereading often are not enough as a main strategy. As a supplement, they do no harm.

There is one further factor that Dunlosky did not evaluate but that is well supported: sleep.

Retrieving instead of rereading

The most important idea first: you learn not only when information goes into your head, but above all when you pull it back out. In research this is called retrieval practice or the testing effect. We explain this in detail in our guide on active recall.

The best-known experiment comes from Roediger and Karpicke (2006). In two controlled laboratory experiments with 120 and 180 students respectively, participants read scientific texts. Some read repeatedly, while others tried to write down the content freely from memory, without feedback. Testing took place after five minutes, two days, or one week.

The interesting part is the pattern. After five minutes, the rereaders were ahead: in Experiment 2 they recalled 83% of the idea units, the retrieval group 71%. After one week the picture reversed completely. Those who had read once and retrieved three times recalled 61%. Those who had read four times recalled only 40%. That is a difference of 21 percentage points.

This is exactly why rereading is so easy to fall for. In the short term it feels better, and it actually is. But the advantage disappears when it matters.

Single experiments can be lucky, so it is worth looking at meta-analyses:

• Rowland (2014) combined 159 independent effects: retrieval practice compared with restudy achieved Hedges' g = 0.50. The differences between studies were large, and the analysis mainly reflects laboratory learning, not regular coursework.
• Adesope, Trevisan, and Sundararajan (2017) analyzed 118 articles with 272 effect sizes and a total of 15,427 participants. Compared with rereading alone, the effect was g = 0.51. The authors stress that the work concerns practice tests with low stakes, and they call for more research in real teaching situations.
• Ruitenburg et al. addressed exactly this gap in October 2026: 53 studies from real school classrooms, 233 effect sizes, 7,303 learners. Retrieval practice there reached g = 0.37. Smaller than in the laboratory, but clearly positive.

Stated honestly: "Answer questions from memory and then check your answers" is very well supported. "This will guarantee you a better grade" is not.

Spreading out instead of cramming

The second high-utility method: spread study time over several days instead of packing everything into one session. This is the spacing effect, the basis of every flashcard app. How spaced repetition works in detail is explained in our guide Spaced repetition explained.

Cepeda et al. published a large synthesis on verbal recall in 2006: 317 experiments from 184 articles. For the comparison of massed versus distributed practice there were 271 comparisons with 14,811 participants. In summary, learners recalled 36.7% after massed practice and 47.3% after distributed practice. This is an average across very different tasks, not a promise for you personally, but a clear signal.

An important distinction that is almost always lost in guides: "spacing" means distributed rather than back-to-back. "Lag" means a longer rather than a shorter gap. These are two different questions.

There is a nice experiment on the second question. Cepeda et al. (2008) had 1,354 people learn 32 facts online, in 26 conditions with different gaps and test dates up to almost a year later. The result: with a fixed test date, performance first improves as the gap between study sessions grows, then declines again beyond a point. The optimal gap produced 64% more free recall than repetition on the same day (d = 1.1). These are relative improvements, not percentage points.

And how large is "optimal"? According to the abstract, roughly 20% of the retention interval when the test is a few weeks away, and about 5% when it is a year away. The gap meant is between the second study session and the test. This does not lead to a magic schedule such as "always review after 1, 3, and 7 days." That is precisely why modern apps calculate the intervals per card rather than following a fixed table.

In real school classes the effect was smaller: Ruitenburg et al. found g = 0.39 for distributed versus massed learning, and no significant difference for "more distributed" versus "less distributed." Spreading out clearly helps. How exactly you spread out matters less in everyday life than the laboratory curves suggest.

Mixing, but deliberately

Interleaving means mixing problem types instead of working through them in blocks. So not twenty tasks of type A followed by twenty of type B, but shuffled together.

The best-known experiment is spectacular and at the same time shaky. Rohrer and Taylor (2007) had students calculate the volumes of different solids, either mixed or blocked. On the delayed test, the mixed group reached 63% and the blocked group 20%. That sounds like a miracle cure. But only 18 students completed all sessions, and another 15 dropped out. With a sample this small, the size of the effect is not reliable.

More robust is the meta-analysis by Brunmair and Richter (2019): 59 studies, 8,466 participants, overall effect g = 0.42 in favor of mixing. The details become interesting:

• For paintings and artist categories, mixing helped strongly (g = 0.67).
• In mathematics it helped moderately (g = 0.34).
• For expository texts, no significant effect was detectable.
• For word material, blocked learning was even better (g = −0.39).

The biggest benefit appeared when categories resemble one another. The practical rule is therefore: mix related, easily confused problem types so that you have to decide each time which method fits. "Switch subjects every half hour" does not follow from this.

Why ask yourself and explain to yourself

Two techniques with moderate utility that fit well with flashcards:

Elaborative interrogation: you explain why a statement is true and connect it with what you already know. According to Dunlosky et al., this tends to work better when you already have prior knowledge. According to our research, there is no robust overall effect size specifically for this method.

Self-explanation: you generate connections, for example "Why is this calculation step allowed?" Here the evidence is stronger. Bisra et al. (2018) analyzed 64 research reports with 5,917 participants and found g = 0.55. One catch: in 47 of the 69 effects it was not reported whether the self-explainers had more time. In the few studies with comparable study time, the effect was g = 0.411. Self-explanation therefore helps, but whether it is better per minute of study than other methods is open.

For flashcards this means: write not only what-cards but also why-cards. Our guide Writing good flashcards shows what this looks like in practice.

Sleep is part of learning

Sleep is not a break from learning. The research overview by Rasch and Born (2013) describes sleep as a phase in which previously learned material is actively stabilized and reorganized, among other things through reactivation and exchange between the hippocampus and the cortex. This is an explanatory model drawn from many studies, not a single experiment.

The numbers:

• Berres and Erdfelder (2021) analyzed 823 effect sizes from 177 articles. Sleep compared with an equally long waking period improved episodic memory by g = 0.44. After correction for possible selective reporting, g = 0.28 remained, still significant.
• Newbury et al. (2021) found: at least one sleepless night after learning worsened memory by g = 0.277 (1,616 healthy adults). Sleep deprivation before learning harmed even more clearly: g = 0.621.

The authors themselves point to publication bias and many small individual studies. And one sleepless night cannot be converted into "this much worse per lost hour." The direction is nevertheless clear: pulling an all-nighter before an exam sabotages exactly what you learned the day before. How to avoid this during the exam phase is described in our guide Study plan for the exam phase.

The learning styles myth

You have certainly heard "I am a visual learner." The idea: if teaching matches your learning style, you learn better.

In 2008 Pashler et al. clearly set out how this would have to be tested. You need groups with different learning styles, random assignment to different teaching formats, the same final test, and a crossover interaction: each group would have to learn better with the format matching its style. The authors found no sufficient evidence that would justify such diagnoses in teaching.

Later work largely confirms this. Aslaksen and Lorås (2018) screened 1,215 records and in the end found only ten studies with matching experiments. Neither visual nor auditory matching produced a significant advantage.

Important, because it is often told wrongly: "All studies find exactly zero" is not correct. A meta-analysis by Clinton-Lisell and Litzinger (2024) found a small advantage for matched instruction across 21 studies (g = 0.31). But only 26% of the learning measures showed the expected pattern, and only five studies met quality standards. The authors themselves consider the advantage too small and too unreliable for broad adoption.

So you are welcome to like diagrams. That is no reason to skip retrieval. Retrieval and spreading out work regardless of which "type" an online quiz has assigned you.

How to apply this tomorrow

1. Read, then close it. After each section, summarize it from memory, then check.
2. Questions instead of highlights. Instead of coloring a sentence yellow, write a question that asks for it. These questions become flashcards.
3. Spread out. Prefer short reviews across several days over one long session on the weekend.
4. Mix similar things. Practice easily confused problem types in a shuffled order.
5. Ask why. For every important fact, write an explanation of why it is true.
6. Sleep. Especially on the night after learning and before the exam.

Where checkst comes in

The two methods with the highest utility, retrieval and spreading out, are exactly what a spaced repetition app automates. checkst shows you each card as a question, you retrieve the answer, rate yourself, and the FSRS algorithm plans when the card comes back. How this works is explained in our guide FSRS explained.

The part that puts most people off, writing the cards, checkst can take off your hands if you wish: the AI suggests cards from your lecture handout, a PDF, or a photo, and you decide card by card which ones to keep. Existing Anki decks can be imported as .apkg. And because learning alone is sometimes tough, there are study groups with shared decks. What checkst cannot do: retrieve for you. You still have to do that yourself. But that is exactly the point.