Loading Motifs…
Let a model predict how likely a learner is to recall each item, and put the items closest to being forgotten first — in review sessions and in the strength display. It helped recall in a classroom; at Duolingo it cut practice return.
+16.5% recallOn an exam 28 days after the semester, over massed review (179 eighth graders, one Spanish course; points or relative not stated). At Duolingo, replacing the old Leitner system cut next-day practice return 7.3%.
Built from published experiments and company reports across several products. Nobody has checked the screens on a dated day, so read each number with the caveat printed beside it.
Let a model predict how likely a learner is to recall each item, and put the items closest to being forgotten first — in review sessions and in the strength display. It helped recall in a classroom; at Duolingo it cut practice return.
Learners notice when a progress display doesn't match what they actually know: Duolingo students complained that strength meters driven by its old Leitner-style system did not reflect what they had learned (Settles and Meeder, Duolingo).
This play also appears inside the Duolingo playbook. You can still use it independently.
Personalised review improved recall a month later in a classroom.
STRONGWhat it does not show One teacher, one school, one course; students were not told the details of the manipulation. Outcome is exam recall, not engagement or voluntary return. Read as the authors' manuscript hosted on Mozer's site; the published version was not compared.
Improving students' long-term knowledge retention through personalized review Lindsey et al. · 2014
Replacing Duolingo's Leitner system cut next-day practice return.
STRONGWhat it does not show This is an adverse measured comparison on one metric, with no significant change on the other two. The authors interpreted the practice drop as favorable and deployed HLR (I1); the experiment does not test that interpretation. Baselines, confidence intervals and per-arm sample sizes are not reported.
A Trainable Spaced Repetition Model for Language Learning Settles and Meeder · 2016
A simpler version of Duolingo's model raised daily retention.
STRONGThe comparison is between two HLR variants, not HLR versus Leitner or versus no spaced review. Baselines, confidence intervals and per-arm sample sizes are not reported. Duolingo's blog presents the same figures as the effect of switching to HLR (S2); the paper is followed here. Measures engagement, not recall or proficiency.
Use the play for this decision alone, or combine it with an existing product strategy after checking for conflicts.
A learner who falls behind faces a growing backlog unless the daily review load is capped.
A model that predicts well on one product's logs may predict poorly on another's.
Judging the change on engagement alone when the point was learning, or the reverse.
Plus everything above, as plain text for your coding agent, with a short instruction on top: check the fit, say what the evidence doesn’t support, and adapt it rather than copy it. Paste it and you get an answer, no prompt to write.
Duolingo's production experiments, a classroom study and an analysis of Duolingo logs · evidence reviewed 25 September 2026.
Explore company-inspired strategies that connect your product, experience, and growth.
Browse the playbooks ↗A Trainable Spaced Repetition Model for Language Learning Settles and Meeder · 2016
Per-word weights made some words decay no matter how often learners practised.
COMPANY-REPORTEDWhat it does not show Reported complaints; prevalence and effect on retention are not quantified. The fix was tested in E1, which combines removing the cause with other changes in model behavior.
A Trainable Spaced Repetition Model for Language Learning Settles and Meeder · 2016
In Duolingo logs, reviews that followed a recall-based schedule showed less forgetting.
SUPPORTINGWhat it does not show Learners chose their own review timing; no schedule was assigned. Read as excerpts of the full text; effect sizes were not recorded.
Enhancing human learning via spaced repetition optimization Tabibian et al. · 2019