The Science | LangArena

Why smarter prep works better.

Retrieve instead of reread, return at the right time, and spend more reps where the gap still matters.

80% vs 33 to 36%
one-week recall in Karpicke & Roediger (2008)
Study snapshot
Repeated retrieval held far more after one week.

One clean example: keep testing after a correct answer and delayed recall stays much higher than stopping early.

Continued retrieval 80%
Stop after first correct 33 to 36%
Karpicke & Roediger (2008) Recall after 1 week
This is the first reason the loop asks for the answer again instead of sending you back to reread.
Retrieval practice
Spacing effect
Adaptive difficulty
IRT calibration
Fisher information
Skill vectors
Gap-first routing
Per-exam tuning
Retrieval practice
Spacing effect
Adaptive difficulty
IRT calibration
Fisher information
Skill vectors
Gap-first routing
Per-exam tuning
The learning loop

The page version of the algorithm.

The algorithm page describes a moving loop. Here is the same idea as one continuous system instead of three copied landing cards.

Signal
Map the skill vector
The diagnostic estimates where you are across separate exam skills, so practice starts from measured gaps instead of a vague overall level.
Separate weaknesses, not one blended score
Choice
Select the highest-value rep
The next item is chosen where uncertainty is still high and the likely score payoff is still meaningful for your target exam.
Use the next rep to learn the most
Return
Revisit with delay, not noise
Practice is meant to return to weak material after useful spacing, rather than burning time on immediate rereads of what already feels easy.
Bring it back when the timing is useful
Each rep can update the estimate, change the next question, and schedule a smarter return.
Selected studies

Built on papers, not slogans.

These are the four findings most relevant to the way LangArena describes its adaptive engine.

Science, 2008
Repeated retrieval beat dropping items after the first correct answer.
Students who kept retrieving learned vocabulary one week later recalled about 80% of the pairs. Students who stopped testing after a single correct recall fell to 33 to 36%.
80% vs 33 to 36% Read paper
Psychological Science, 2008
The best review gap depends on when you need the knowledge.
Cepeda and colleagues found that the optimal spacing window shrinks as a share of the final delay, from about 20 to 40% of a one-week target to 5 to 10% of a one-year target.
20 to 40%, then 5 to 10% Read paper
J Urol, 2010
Adaptive spacing raised learning efficiency without lowering outcomes.
In a randomized trial, adaptive spaced education achieved comparable end-of-course scores while using fewer practiced items, increasing learning efficiency by 38%.
+38% efficiency Read paper
Computers & Education, 2011
Computer adaptive practice works best near the edge of success.
Klinkenberg and colleagues ran adaptive arithmetic practice at a mean success probability of 0.75 and reported better precision, high validity, and strong engagement across millions of responses.
Target success: 75% Read paper
Research takeaways

The repeated pattern is active, timed, selective practice.

Across the studies used on this page, the winning approach is not more content. It is better retrieval, better timing, and better targeting.

Delayed retention
80%
Recall one week later when repeated retrieval stayed in the loop.
Lift vs early drop
>150%
Retention gain over stopping repeated testing after the first success.
Adaptive efficiency
38%
Learning efficiency gain in adaptive spaced education.
Challenge zone
0.75
Mean success probability used in adaptive computer practice.
Retrieval over rereading

Exposure feels fluent. Retrieval actually sticks.

The strongest pattern in the evidence is not subtle: once learners can answer correctly, continuing to retrieve that knowledge leads to much better delayed retention than just looking at it again or dropping it too early.

This is why the algorithm page talks about a live loop instead of a content library. The point is not to show more material. The point is to keep pulling on the weakest, most informative memory traces.

Karpicke & Roediger, 2008 Roediger & Karpicke, 2006
Delayed recall after learning
Keep retrieving vs stop after one correct answer
Repeated retrieval kept in the loop 80%
Dropped from repeated testing after first success 35%
Average comparison from the four-condition vocabulary experiment reported in Science.
Source: Karpicke & Roediger, 2008
Adaptive sequencing

More efficient means fewer wasted reps.

Adaptive practice is worth doing when it changes the ratio between effort and gain. In Kerfoot's randomized trial, the adaptive version matched the end-course test with fewer practiced items and a 38% lift in learning efficiency.

That is the product logic behind gap-first routing. If inference, pacing, or delivery is the real bottleneck, the loop should spend reps there instead of repeating whatever is already stable.

Kerfoot, 2010 Warner et al., 2020
Normalized learning efficiency
Adaptive spaced education vs fixed spaced education
Adaptive spaced system 138
Fixed spaced system 100
Same end-course test, fewer practiced items
Index derived from the reported 38% efficiency improvement. Fixed condition = 100.
Source: Kerfoot, 2010
Spacing tied to the test date

There is no single perfect interval.

The right revisit time depends on the retention horizon. If the exam is close, the review should return soon. If the exam is far away, the gap can stretch. Research keeps showing that cramming everything into one block is an inefficient default.

This is also why a smart prep system should not treat every missed item the same way. Some gaps need immediate correction. Others need a later return at a more useful delay.

Cepeda et al., 2008 Cepeda et al., 2006
Optimal review window
Derived from the ratios reported by Cepeda et al.
If your exam is 1 week away Best review gap: about 20 to 40% of the delay
Derived window: roughly 1 to 3 days later
If your exam is 1 year away Best review gap: about 5 to 10% of the delay
Derived window: roughly 18 to 36 days later
0% 25% 50% 75% 100%
Source: Cepeda et al., 2008
How LangArena applies it

The research translated into product decisions.

This is the bridge from the papers above to the claims on the algorithm page.

IRT calibration
Question difficulty and learner ability update from responses, so the next rep is chosen from evidence instead of a static sequence.
Multi-skill telemetry
Reading, listening, speaking, grammar, vocabulary, and inference can move separately, which is closer to how exam performance actually breaks down.
Gap-first routing
The engine prioritizes the weakest dimensions with the biggest expected score return, rather than padding already-stable skills for comfort.
Per-exam calibration
The target profile is different for TOEFL, IELTS, TOEIC, Cambridge, PTE, Cyprus Greek, and EIKEN, so the route shifts with the exam.
Ready?

See your own gaps, not just the theory.

Start the diagnostic and watch the adaptive loop map what to train next.

Run free diagnostic
No credit card. Results in minutes.