Long-Term Learning: Mechanisms, Evidence, and Failure Modes

A century of memory research converges on an uncomfortable conclusion: the study behaviors that feel most productive — rereading, highlighting, massed practice, smooth error-free lessons — are optimized for short-term performance and do almost nothing for long-term retention. The two interventions with the strongest, most replicated evidence are retrieval practice (being forced to recall from memory) and spaced practice (distributing that recall over time), and both feel worse to the learner while they are happening. This dissociation between performance during learning and durable learning is the single most important fact for anyone building an education system, human or AI. It explains why learners systematically choose bad strategies, why classroom structures resist the effective ones, and why an AI tutor's core loop must be built around a per-concept memory model and a retrieval scheduler rather than around content delivery. This note explains the underlying memory mechanisms, walks through the experimental evidence and how the field arrived at it, corrects several numbers that are routinely misquoted (including Bloom's 2-sigma and the DKT benchmark results), covers the algorithms that operationalize the science (SM-2, half-life regression, FSRS, Bayesian and deep knowledge tracing), and catalogs the failure modes — what not to build. Every quantitative claim below has been checked against the primary paper.

Background: Learning Is Not Performance

The foundational distinction, formalized by Robert Bjork, is between performance (how well you can do something right now) and learning (a durable change in what you can do later, in other contexts). The two are not just different — they are experimentally dissociable: conditions that raise performance during practice often lower long-term retention, conditions that suppress practice performance often raise it, and learning can occur with no visible performance change at all (Soderstrom & Bjork, 2015).

Bjork's "new theory of disuse" (Bjork & Bjork, 1992) gives the mechanism a two-variable model. Every memory has:

• Storage strength — how well-entrenched the memory is. The theory assumes it is never lost; it only accumulates, and it grows most when the memory is retrieved while it is hard to retrieve.
• Retrieval strength — how accessible the memory is right now. It rises fast with recent exposure and decays with time and interference.

The critical interaction, stated directly in the theory: when an item with low retrieval strength is successfully retrieved, it receives a greater increment in storage strength than an item retrieved at high retrieval strength. Recalling something you were about to forget strengthens it far more than recalling something you saw five minutes ago. This one asymmetry generates most of the field's headline results: spacing works because the gap lets retrieval strength decay before the next retrieval; massed practice fails because each repetition arrives while retrieval strength is still high and therefore deposits almost nothing; and learners misjudge their own learning because they can only feel retrieval strength (current fluency), not storage strength (durability).

The forgetting curve

Hermann Ebbinghaus (1885) produced the first quantitative description of forgetting. He memorized lists of nonsense syllables (to eliminate prior-knowledge effects), waited, and then measured the savings — how much less time relearning took compared to original learning. His data:

A direct modern replication (Murre & Dros, 2015) reproduced the shape of the curve closely from 20 minutes through 6 days using Ebbinghaus's own method, and additionally found a discontinuity — a jump in savings around the 1-day point — consistent with sleep-related consolidation. Sleep is not a footnote here: Mazza et al. (2016) had learners study foreign vocabulary in two sessions 12 hours apart and found that placing sleep between the sessions halved the practice needed to relearn and produced better retention at both 1 week and 6 months. A scheduler that never spaces across a night is leaving consolidation on the table.

Two properties of the curve matter. First, forgetting is steep initially and flattens — roughly exponential in form, so it is well approximated by retention functions like R(t) = exp(-t/S) where S is a stability parameter. Second, and less famous: each relearning episode flattens the curve. Stability S grows with every successful spaced retrieval, so after a handful of well-timed retrievals a memory becomes effectively permanent. Every spaced-repetition scheduler from Leitner boxes to FSRS is a strategy for estimating S per item and scheduling the next retrieval just before R(t) falls below an acceptable threshold.

The Two Techniques That Survived a Century of Scrutiny

Dunlosky et al. (2013) evaluated ten popular learning techniques against the full evidence base, grading each on whether benefits generalize across learners, materials, and retention intervals (their Table 4). Only two earned the top rating, and the ranking is worth internalizing because the bottom of the list is what students actually do:

Retrieval practice: the testing effect

The modern anchor experiment is Roediger & Karpicke (2006, Experiment 2). Students learned prose passages under three regimes — four study periods (SSSS), three study periods plus one free-recall test (SSST), or one study period plus three free-recall tests (STTT) — with no feedback on any test. The result inverts depending on when you measure:

Massed studying wins the quiz you take today (d = 1.22 at 5 minutes); repeated retrieval wins the exam you take next week (d = 1.26). Proportional forgetting over the week: 52% for the repeated-study group, 14% for the repeated-testing group. Crucially, the students' own predictions tracked the 5-minute result — mean judgments of one-week retention were 4.8 (SSSS) vs 4.0 (STTT) on a 7-point scale; the most confident group retained the least. Retrieval is not measurement of learning; the act of retrieval is itself a memory modifier, and a stronger one than re-exposure.

The effect is large and replicable at meta-analytic scale: Rowland (2014) found g = 0.50 (CI [0.42, 0.58]) across 61 studies / 159 effects; Adesope, Trevisan & Sundararajan (2017) found g = 0.61 (CI [0.58, 0.65]) across 272 effects from 118 articles, with classroom settings at g = 0.67. Karpicke & Blunt (2011, Science) showed retrieval practice beat elaborative concept mapping (0.67 vs 0.45 on a one-week test, d = 1.50) — and the advantage held on inference questions and even when the final test was concept mapping (d = 1.01), countering the intuition that retrieval only drills rote facts. In that study 84% of students performed better after retrieval practice while 75% predicted the opposite. Real classroom deployments need nothing fancy: in 6th-grade social studies, low-stakes quizzing lifted chapter exams from 81% to 94% and end-of-semester exams from 67% to 79% (Roediger, Agarwal, McDaniel & McDermott, 2011); in 8th-grade science, quizzing produced 13–25% gains on summative exams (McDaniel et al., 2011). Closed-book "brain dumps" and mini-quizzes qualify (Agarwal & Bain, 2019).

Four boundary conditions matter for system design:

• Feedback roughly doubles the effect. Rowland's moderator analysis: g = 0.73 with feedback vs 0.39 without. Feedback corrects errors and turns failed retrievals into learning events — and delayed feedback outperformed immediate feedback in Butler & Roediger (2008): one-week cued recall of 0.56 (delayed) vs 0.45 (immediate) vs 0.31 (no feedback).
• Format matters less than folklore says. The "effortful formats beat recognition" intuition is not consistently supported: Rowland found cued recall (g = 0.61) above free recall and recognition (both g = 0.29) in the full dataset; Adesope found multiple-choice (0.70) above short-answer (0.48); Smith & Karpicke (2014) found no format difference across four experiments. Retrieval success plus feedback matters more than the response format.
• Difficulty of successful retrieval is dose-dependent. Pyc & Rawson (2009): the harder a successful retrieval was (longer intervals between practice trials), the better the final retention — while over-drilling to high within-session criteria showed diminishing returns. This is the retrieval-effort mechanism behind spacing, stated as a scheduling parameter.
• Transfer is real but bounded. Pan & Rickard's (2018) meta-analysis of transfer of test-enhanced learning: d = 0.40 overall, strongest to different test formats and application/inference questions, weakest to untested sibling material seen during study — and often no positive transfer when no favorable moderator is present. A tutor cannot assume that drilling one item strengthens its untested neighbors; coverage has to be explicit.

Two results turn retrieval practice into policy. Karpicke & Roediger (2008, Science) had students learn Swahili–English pairs and then either kept testing items after the first correct recall or dropped them. Learning curves were identical; one week later, the kept-in-testing conditions sat at ~80% recall and the dropped conditions at 33–36% — d = 4.03, with non-overlapping distributions, while repeated studying after the first correct recall added nothing. Dropping an item from retrieval the moment it is "known" is the single most destructive scheduling policy available. And Rawson & Dunlosky (2011; 533 students, 1–4-month retention) turned the compound intervention — successive relearning — into a dosage prescription: retrieve each item to an initial criterion of ~3 correct recalls, then relearn it to one correct recall in ~3 widely spaced sessions; additional relearning sessions past three were "not time well spent." That loop is exactly what a tutoring system should automate.

One more free lunch: the pretesting effect. Attempting to answer before studying — even when the attempt is guaranteed to fail — improves subsequent learning. Richland, Kornell & Kao (2009): pretest accuracy was 5%, yet pretested material was learned at 75% vs 56% for extended-study controls (d = 1.1). Kornell, Hays & Bjork (2009) got the same result with unanswerable fictional trivia (0.41 vs 0.31, d = 0.58) with answer-exposure time equated. A tutor should pose the question before the explanation; wrong guesses are not wasted time.

Spaced practice: the spacing effect and the ridgeline

Cepeda, Pashler, Vul, Wixted & Rohrer (2006) meta-analyzed 317 experiments (839 assessments from 184 articles): distributed practice reliably beats massed practice at every retention interval tested, and — the key structural finding — the optimal gap between study sessions grows with how long you need to remember. The follow-up (Cepeda et al., 2008), a 1,354-person study crossing gaps of 0–105 days with retention intervals (RIs) of 7–350 days, mapped the "temporal ridgeline." The observed optima for recall:

Three design-relevant facts fall out of this surface:

• It is an inverted U: too-short gaps behave like massing (retrieval strength never decays, so little is deposited); too-long gaps mean the memory is gone before the second session. Overall, studying at the optimal gap rather than a zero-day gap improved final recall by 64% (d = 1.1) with total study time held constant.
• The optimal gap grows with the retention interval, but as a shrinking fraction of it — the paper's own summary: from about 20–40% of a 1-week test delay down to about 5–10% of a 1-year delay. For year-scale retention, the observed optimum was a roughly three-week gap.
• The penalty is asymmetric: "as gap increases, accuracy increases steeply and then declines much more gradually" — a too-long gap costs much less than a too-short one. A scheduler under uncertainty should err on the long side.

The practical rule: re-space after each successful retrieval, since each retrieval raises stability and pushes the next optimal gap outward — expanding intervals over successive retrievals. (Whether expanding schedules beat equal-interval schedules for a fixed number of retrievals is genuinely unresolved: Karpicke & Roediger (2007) found expanding better at 10 minutes but equal spacing better at 2 days, and Cepeda et al.'s (2006) review of 18 such studies calls the evidence tentative. The load-bearing variable is that gaps exist and are roughly ridgeline-sized, not their precise shape.)

The effect holds in young children — Seabrook, Brown & Solity (2005) found distributed phonics teaching (three 2-minute sessions/day) produced over six times the improvement of a single 6-minute session in 5-year-olds (8.3 vs 1.3 points) — and it survives contact with real classrooms at system scale: Lindsey, Shroyer, Pashler & Mozer (2014) deployed a personalized spaced-review scheduler (a per-student, per-item memory model) in a semester-long middle-school Spanish course and improved cumulative-exam retention by 16.5% over massed review and 10.0% over one-size-fits-all spacing. That study is the existence proof for the central engineering claim of this note: modeling individual forgetting and scheduling against it beats both no spacing and generic spacing, in a real course, on a delayed exam.

Desirable Difficulties and the Metacognitive Illusion

Bjork (1994) named the umbrella principle desirable difficulties: manipulations that slow or impair acquisition while enhancing long-term retention and transfer. The canonical family: spacing, interleaving (contextual interference), varying practice conditions, reducing feedback frequency, and using tests rather than presentations as learning events — with generation (producing answers rather than reading them; Slamecka & Graf, 1978) added in later formulations.

Interleaving

Mixing problem types instead of blocking them is the sharpest demonstration that practice performance and learning move in opposite directions:

Note the 2015 pattern: interleaved practice was "near immunity against forgetting" — a 30-fold increase in test delay cost 6 points (80→74) for interleavers and 22 points (64→42) for blockers. The interleaving advantage is substantially a forgetting-resistance advantage, which is why short-delay tests understate it.

The mechanism is discriminative contrast, not spacing in disguise: interleaving forces the learner to select a strategy on every problem (the skill exams and life actually demand), whereas blocking lets them re-execute a known procedure. Two results pin this down. Birnbaum, Kornell, Bjork & Bjork (2013): inserting 10-second gaps between category exemplars — adding spacing while destroying juxtaposition — eliminated the interleaving benefit. Kang & Pashler (2012): interleaving beat a blocked-but-equally-spaced control (68% vs 61%, d = 0.78) with temporal spacing equated.

But interleaving is the most boundary-conditioned of the desirable difficulties, and an honest summary needs the moderator map. Brunmair & Richter's (2019) meta-analysis (59 studies, 238 effects) found overall g = 0.42, decomposing to:

Interleaving helps when categories are mutually confusable (high between-category similarity) and hurts when items are unrelated — consistent with Carvalho & Goldstone (2014), who showed the advantage flips to blocking for low-similarity categories. Rohrer's own papers add the practical caveats: keep a small initial block when a skill is first introduced, expect interleaved assignments to take more time, and provide corrective feedback, since interleaved practice produces more errors.

The metacognitive illusion

The dark side of every desirable difficulty is that learners judge learning by fluency — how easily material is processed right now — which is a readout of retrieval strength, not storage strength. The cleanest demonstration is Kornell & Bjork (2008): learning painters' styles from exemplars, spacing/interleaving beat massing decisively (d ≈ 1.0–1.3), yet — judging after taking the final test — 78% of participants performed better with spacing while 78% rated massing as good or better. The illusion survives direct experience of the outcome. Students also misuse the one effective technique they do adopt: 91% of surveyed undergraduates self-test, but only 18% treat testing as a learning event rather than a knowledge check (Kornell & Bjork, 2007).

This has a hard product consequence: a system that optimizes for learner-perceived progress or in-session success will be selected against the interventions that work. Any effective system must either make the difficulty legible ("this is supposed to feel hard; here is why") or absorb the strategy choice into the system so the learner never has to choose against their intuition.

Reading the Effect-Size Evidence Without Fooling Yourself

Education research communicates in Cohen's d (difference in means divided by pooled standard deviation). Hattie's synthesis puts the average of all studied educational interventions around d = 0.40, a useful "hinge": an intervention at d = 0.40 is merely average, not good. Here is the hierarchy relevant to tutoring, with the trust caveats that summaries usually omit:

Read the ITS rows top to bottom and a pattern appears that generalizes across this whole literature: effects shrink as studies move from lab to field, from experimenter-made tests to standardized tests, and from small deployments to scale (0.76 → 0.66 → 0.42 → 0.37 → 0.20). None of these numbers is "the" effect of tutoring software; the funnel itself is the finding.

Three corrections to the folklore version of this table:

Bloom's 2-sigma is a ceiling for a bundle, not a fact about tutoring. Bloom (1984) reported that tutored students scored two standard deviations above conventional 30:1 classroom instruction — the average tutored student above 98% of the control class. But the underlying data are two dissertations from his own group (Anania; Burke): grades 4–8, single well-defined units (probability, cartography), three weeks of instruction — and the tutoring condition included mastery learning's formative-test-and-correctives loop; Bloom's own Figure 2 labels the 2.0σ bar "tutoring + mastery learning," and neither dissertation evaluated tutoring alone. Mastery learning by itself was ~1σ in the same studies, so tutoring's increment over mastery procedures was roughly 0.8σ (Kulik & Fletcher, 2016). Modern evidence puts replicated human tutoring far lower: VanLehn (2011) found d = 0.79 across controlled studies, and Nickow et al.'s meta-analysis of 96 randomized K-12 tutoring trials found a pooled d = 0.37 (strongest for teacher and paraprofessional tutors and in early grades). The right reading: 2.0 is what individualization + mastery + immediate feedback achieved under ideal conditions on aligned tests — the headroom target, not a benchmark any deployed system has hit.

VanLehn's granularity result is a plateau, not a ladder. The interaction granularity hypothesis predicted finer-grained feedback → more learning, all the way down. The data said otherwise: moving from answer-based CAI (d ≈ 0.30, the Kulik & Kulik 1991 benchmark VanLehn used) to step-based tutoring (d = 0.76) is a huge win, but substep-based tutoring came in at d = 0.40 — nominally worse, and certainly no better. Step-based ITS matched human tutors (0.76 vs 0.79). The design implication is precise: give feedback and hints at the level of solution steps — not only at final answers, and not micro-managing below the step. Beyond that granularity, invest elsewhere.

Hattie's rankings are directional, not precise. Visible Learning aggregates 800+ meta-analyses covering wildly different outcome measures, populations, and designs. The numbers move between editions (formative evaluation went from 0.90 to 0.48; grade retention from −0.16 to −0.32), and methodologists have documented outright calculation errors — Bergeron & Rivard (2017) show Hattie's "common language effect" values, supposedly probabilities, range from −49% to 219%. Use the hierarchy to sort interventions into tiers — clearly strong / average / harmful — and never to claim "X is exactly d = 0.73."

Feedback deserves its own warning label, since it is both powerful and dangerous. Kluger & DeNisi's (1996) classic meta-analysis (607 effect sizes): average d = 0.41, but 38% of feedback interventions reduced performance. Hattie & Timperley (2007) explain the variance with four feedback levels: task, process, self-regulation, and self. Process- and self-regulation-level feedback ("your error pattern suggests you're forgetting to check the sign when distributing") drives learning; self-level feedback (praise — "you're so smart!") is rarely effective and can backfire. Wisniewski, Zierer & Hattie (2020, 435 studies) confirm: high-information feedback works best, reinforcement/punishment worst. This is the strongest argument for treating feedback generation as a designed, evaluated component rather than a place where an LLM freestyles.

What about LLM tutors specifically? The first credible RCTs (2023–2026) split cleanly along this note's central axis. Kestin et al. (2025): a carefully scaffolded GPT-4 physics tutor at Harvard beat in-class active learning with effect sizes of 0.73–1.3 SD — on immediate post-tests. A World Bank RCT in Nigeria (800 students, six weeks of after-school GPT-4-assisted English) found +0.23 SD on English outcomes at unusually low cost. But Bastani et al. (2025, high school math): unrestricted GPT-4 access improved practice performance by 48% and then reduced unaided exam performance by 17% — students used the model as an answer engine and learned less than the no-AI control — while a version with tutor-style guardrails (hints, no direct answers) eliminated the harm. And no LLM-tutor RCT to date reports delayed retention outcomes, which by the performance/learning dissociation means even the positive results are so far performance-side evidence. The technology amplifies whichever loop you build: scaffolded retrieval-and-hints, or frictionless answer extraction.

Mastery Learning: The Mechanism and Why Classrooms Dropped It

Bloom's Learning for Mastery (1968), building on Carroll's (1963) model of school learning, inverts the classroom constant: conventional instruction holds time constant and lets achievement vary; mastery learning holds achievement constant and lets time vary — aptitude is re-conceived as the rate at which a student learns, not a ceiling on what they can learn. The loop: teach a unit → formative assessment → students below the mastery threshold get corrective instruction (a different explanation, not just more of the same) → students above it get enrichment → parallel reassessment → progress once mastery is demonstrated. The theoretical justification is cumulative: most curricula are prerequisite chains, so a student who "passes" a unit at 70% carries a 30% hole into every subsequent unit that builds on it, and gaps compound into what looks like inability but is actually accumulated prerequisite debt.

The evidence: Kulik, Kulik & Bangert-Drowns (1990), across 108 studies (103 with end-of- instruction exams), found d = 0.52 overall, rising to 0.64 where the mastery threshold was strict (91–100%), with retention follow-ups averaging d = 0.71 at ~8 weeks. Honest caveats, because mastery learning is also the clearest case study in evaluation hygiene: effects on experimenter-made tests averaged 0.57 while effects on standardized tests averaged 0.29 — and for Bloom-style LFM programs specifically, the five studies with standardized tests found essentially nothing (≈ 0.10). Slavin's (1987) best-evidence synthesis, restricted to studies with standardized measures and equal time, found a median near 0.25 and argued the robust effects are test-alignment artifacts. Time cost is disputed the same way: Kulik et al. measured ~4% extra time in typical implementations, while Guskey's practitioner literature estimates correctives add 10–20% to initial units. The defensible claim: mastery learning reliably improves performance on the material actually taught and assessed, with the size of the "true" gain depending on how far the outcome measure sits from the taught unit — which is precisely Pan & Rickard's transfer boundary appearing at curriculum scale.

Why did classrooms drop it? Bloom's group-based LFM keeps the class together, which means early-masters wait while correctives run — and the individualized variants (Keller's PSI) that let students move at their own pace collapse under the bookkeeping: managing thirty students at different curriculum positions, authoring parallel assessment forms and multiple corrective explanations per unit, and fighting the grading system (Guskey, 2010, describes the machinery and its costs; the decline itself is documented across the implementation literature). This is the important asymmetry for AI tutoring: every reason mastery learning failed in classrooms is a bookkeeping problem, and bookkeeping is what software is good at. Per-student pacing, unlimited parallel assessment variants, and alternate explanations on demand are exactly the parts an LLM-based system makes cheap.

Two design rules follow. First, mastery must mean more than a percentage: a student can hit 80% via guessing, pattern-matching, or memorized procedures ("fragile knowledge"); demonstrated mastery should include explanation or transfer, and low-confidence correct answers should trigger an "explain your reasoning" probe (prompted self-explanation independently carries g ≈ 0.55 — Bisra et al., 2018). Second, remediation must traverse the prerequisite graph backward — if a student fails fraction addition, the fix is often a decayed prerequisite (equivalent fractions), not more fraction-addition drills.

The Machinery: Student Modeling and Retrieval Scheduling

Everything above becomes implementable through two components: a model of what the student knows (knowledge tracing) and a policy for when to bring each item back (scheduling).

Bayesian Knowledge Tracing (BKT)

Corbett & Anderson (1994) is still the reference student model. Each skill is a two-state hidden Markov model — known or not known — observed through noisy correct/incorrect responses. Four parameters per skill:

• P_L0 — probability the skill is known before practice starts
• P_T — probability of transitioning to known after a practice opportunity
• P_guess — probability of answering correctly while not knowing
• P_slip — probability of answering incorrectly while knowing

# state per (student, skill): L = current belief that skill is known L = P_L0 on each attempt with observed response: # 1. Bayesian update on the evidence if correct: posterior = L*(1 - P_slip) / (L*(1 - P_slip) + (1 - L)*P_guess) else: posterior = L*P_slip / (L*P_slip + (1 - L)*(1 - P_guess)) # 2. Learning transition (student may have just learned it) L = posterior + (1 - posterior) * P_T # prediction for the next attempt: P(correct next) = L*(1 - P_slip) + (1 - L)*P_guess mastery declared when L >= 0.95 # the Cognitive Tutor's criterion

(The 1994 paper states the update as Bayesian inference plus the transition equation; the explicit posterior formulas above are the standard derivation, made explicit in later presentations.) Parameters are fit per skill from historical data (expectation-maximization or grid search). BKT's virtues: interpretable, works with sparse data, and its mastery estimate directly drives pedagogy. Its known problems: skills are assumed independent (false — that is what prerequisite graphs are for); no forgetting — L never decreases, so vanilla BKT literally cannot represent the forgetting curve and must be extended with a decay term for any long-term system; and the identifiability problem (Beck & Chang, 2007) — an infinite family of parameter sets makes identical performance predictions while implying different knowledge states, which is why practitioners bound the parameters (the common "bounded BKT" convention, P_guess ≤ 0.3 and P_slip ≤ 0.1, comes from this later degeneracy work — Baker, Corbett & Aleven 2008 lineage — not from the original paper). Baker et al. (2008) also improved accuracy by making slip/guess contextual, using response time and help requests as evidence.

The logistic family: IRT, PFA, and Elo

The other classical lineage models P(correct) directly as a logistic function rather than tracking a hidden knowledge state. Item Response Theory (Rasch, 1960; Lord, 1980) is the psychometric backbone: P(correct) = σ(θ_student − b_item), one ability parameter per student, one difficulty per item — the model behind adaptive testing. Performance Factors Analysis (Pavlik, Cen & Koedinger, 2009) extends it with per-skill counts of prior successes and failures, giving an interpretable, incrementally updatable student model with no EM fitting. Elo-family ratings (Klinkenberg et al., 2011 — the Math Garden system; Pelánek, 2016) are the production workhorse: student ability and item difficulty updated online after every response, exactly like chess ratings, which is the cheapest way to run the difficulty-targeting controller described under ZPD below. These matter because they are the actual baselines that modern deep models get compared against — and, as it turned out, largely tie.

Deep Knowledge Tracing and the hype correction

Piech et al. (2015) replaced the per-skill HMM with a single LSTM over the student's full interaction sequence, reporting AUC 0.86 on ASSISTments vs 0.67–0.69 for BKT — "a 25% gain over the previous best" — and launched a subfield. The correction arrived within a year and is a case study in benchmark hygiene. Xiong et al. (2016) found the benchmark dataset contained 23.6% duplicated rows plus scaffolding items the baselines had excluded; on cleaned data, DKT scored 0.73–0.75 — statistically tied with plain logistic PFA at 0.73. Khajah, Lindsey & Mozer (2016) showed classical models given the same affordances close the rest of the gap: BKT with a forgetting term reached 0.83 vs DKT's 0.86, and BKT variants with exercise-indexed difficulty beat DKT outright (0.90); their verdict was "indistinguishable on average." Wilson et al. (2016) found hierarchical IRT matched or beat DKT on every dataset. Gervet et al. (2020), across nine datasets: well-featured logistic regression leads on moderate-size data, deep models lead on the largest datasets or where fine temporal structure matters, and BKT-family models lag both. Later architectures — DKVMN's key-value memories (Zhang et al., 2017), attention-based SAKT (Pandey & Karypis, 2019) and AKT (Ghosh et al., 2020) — improved benchmarks further, but a decade in, essentially no production tutor runs deep knowledge tracing. The reasons are structural, not fashion: deep models are cold-start hungry, and — decisive for tutoring — they predict response correctness without exposing which skill is deficient and why, so they cannot justify pedagogical decisions to students or teachers. The pragmatic architecture: interpretable per-skill state (BKT-with-forgetting or the logistic family) as the system of record, optionally sharpened by a sequence model, with the emerging frontier being LLM-based assessment of open-ended student explanations feeding evidence into that structured state.

Scheduling: from Leitner boxes to FSRS

Scheduling algorithms operationalize the ridgeline: estimate each item's memory stability and schedule retrieval just before predicted recall drops below a target (typically 90%).

• Leitner (1972): items move through boxes with increasing review intervals; failure demotes the item to box one. A step-function approximation of expanding intervals — trivially simple, surprisingly effective.
• SM-2 (Wozniak, 1987; the basis of classic Anki): per-item easiness factor and multiplicative interval growth:

# state per item: n (consecutive successes), EF (easiness, init 2.5), I (interval, days) # input after each review: quality q in 0..5 if q < 3: # failed recall n = 0 I = 1 # relearn from short interval; EF unchanged else: n += 1 if n == 1: I = 1 elif n == 2: I = 6 else: I = round(I * EF) EF = max(1.3, EF + 0.1 - (5 - q) * (0.08 + (5 - q) * 0.02)) schedule next review at today + I # (original SM-2 also re-drills any item scoring < 4 later the same day until it reaches >= 4)

Note what SM-2 encodes: intervals expand after success (stability grew), failure resets toward short intervals, and per-item difficulty modulates growth. Its weakness is that the update rules are hand-tuned constants, not fit to data.

• Half-Life Regression (Settles & Meeder, 2016 — Duolingo): model recall as p = 2^(-Δ/h) where Δ is time since last practice and half-life h = 2^(θ·x) is regressed from features (correct/incorrect counts, item difficulty) over 12.9M review traces. Cut recall- prediction error by ~45% vs a tuned Leitner baseline (MAE 0.128 vs 0.235). The step from hand-tuned rules to a learned forgetting model.
• Optimal-control formulations: Tabibian et al. (2019, PNAS) framed spaced repetition as stochastic optimal control and derived MEMORIZE, a provably optimal policy in which review intensity is proportional to the item's current forgetting rate — the theoretical bridge between regression models and deployed schedulers, validated on Duolingo data.
• FSRS (derived from Ye, Su & Cao, KDD 2022; Anki's modern scheduler since 23.10): implements the three-component model directly — per-item Difficulty, Stability, and Retrievability, with a power-law forgetting curve (FSRS v4: R(t) = (1 + t/(9S))^(-1); later versions generalize the exponent) — and fits its 17–21 parameters to the individual learner's review history. On the open srs-benchmark (~10,000 Anki users, ~727M reviews), FSRS predicts recall better than SM-2 for 99.6% of users, which translates to fewer reviews for the same retention. This is the current practical state of the art for the scheduling half of the problem — and Lindsey et al. (2014) is the classroom RCT showing this model class pays off on real course outcomes.

A tutor-scale scheduler adds constraints flashcard apps ignore: retrieval events should be interleaved across confusable topics (a free desirable difficulty — subject to the Brunmair & Richter similarity boundary), varied in format with feedback attached (format itself matters less than feedback presence), never dropped at first success (Karpicke & Roediger 2008: d = 4.03 for staying in the queue; Rawson & Dunlosky's 3-recalls-then-3-relearnings rule is the dosage), biased long under uncertainty (the ridgeline's asymmetric penalty), and prioritized by prerequisite centrality — letting a load-bearing concept decay costs more than letting a leaf concept decay.

Targeting difficulty: the Zone of Proximal Development, operationalized

Vygotsky's ZPD (1978) — the band between what a learner can do alone and what they can do with help — is the difficulty-targeting principle. Murray & Arroyo (2002) operationalized it for tutoring systems as a target band of help-usage: adapt task difficulty so the student needs some, but bounded, assistance. Concretely: (1) dynamic assessment — pose a problem at the edge of estimated ability; if the student fails, give a graded hint and observe whether it unlocks the solution (success-with-one-hint ≈ inside the ZPD; failure-despite-hints ≈ beyond it); (2) success-rate control — regulate difficulty to hold success around 80–90% (an Elo-style online rating makes this a one-line controller), high enough to build self-efficacy, low enough that retrieval stays effortful; (3) scaffold fading — support must be withdrawn as competence grows, or the support becomes a dependency. A workable hint ladder (a synthesis, not a validated protocol): first error → targeted hint at the failing step; second → conceptual hint; third → worked example; fourth → drop back to the prerequisite concept. The key insight: the ZPD is not a property of the student but of the student-task pair — it must be re-estimated per skill, not assumed global. LLMs are genuinely strong here: generating hints at graduated specificity, tailored to the student's actual error, was hand-authored content in classic ITS and is now cheap.

Sequencing Instruction: Cognitive Load and the Guidance Continuum

Cognitive Load Theory (Sweller, 1988) supplies the constraint that shapes how to teach at each expertise level. Working memory holds roughly four chunks (3–5; Cowan, 2001) for seconds without rehearsal; long-term memory is effectively unlimited; learning is the construction of schemas in long-term memory that let bigger chunks pass through the working-memory bottleneck. Load splits into intrinsic (inherent element interactivity of the material), extraneous (imposed by bad presentation — the part instruction can eliminate), and germane (the productive effort of schema construction). The replicated effects and their design translations:

The same logic frames the direct-instruction-versus-discovery debate. Kirschner, Sweller & Clark (2006) argue from the working-memory bottleneck that minimally-guided instruction fails novices: searching a problem space consumes the entire working-memory budget, leaving nothing for schema construction. The evidence supports them — for unguided discovery with novices. The expertise reversal effect supplies the other half: as schemas form, guidance becomes redundant and problem-solving becomes the better use of time. And there is a designed exception in between: productive failure (Sinha & Kapur, 2021; meta-analysis of 166 comparisons) — a structured problem-solving attempt on contrasting cases before instruction — beat instruction-first on conceptual knowledge and transfer (g = 0.36, up to ≈ 0.58 for high-fidelity designs) without hurting procedural knowledge, though the effect reverses for young learners (grades 2–5). Note that this is the pretesting effect scaled up to whole lessons: a failed generation attempt prepares the mind for the explanation that follows. So "direct instruction vs. inquiry" is a false dichotomy; the real variables are where the learner sits on the novice-expert continuum for this specific concept and whether the struggle is designed and bounded. The correct policy is a per-concept slide from full guidance (direct instruction, worked examples) through completion problems to independent, interleaved practice — optionally prefixed, for older students, by a short bounded problem-solving attempt. Chi & Wylie's (2014) ICAP framework gives the activity-selection heuristic within any of these stages: learning outcomes ordered Interactive > Constructive > Active > Passive — prefer formats where the student generates or co-constructs over formats where they select or receive.

Motivation: The Part That Determines Whether Any of This Runs

The cognitive interventions only work if the student keeps showing up, and spaced retrieval — the most effective intervention — is also the least pleasant. Motivation science is what closes that gap, and its headline results are mostly warnings.

Self-Determination Theory (Deci & Ryan) identifies three needs that sustain intrinsic motivation: autonomy (acting volitionally — controlling language and surveillance measurably reduce engagement, depth of processing, and persistence; Vansteenkiste et al., 2004), competence (the experience of mastering optimal challenges — which the 80–90% success-rate policy delivers directly), and relatedness (feeling connected — the channel through which tutor persona and long-term relationship plausibly matter).

The gamification trap. Deci, Koestner & Ryan's (1999) meta-analysis of 128 experiments: expected tangible rewards undermine intrinsic motivation for interesting tasks (engagement-contingent d = −0.40; the overjustification effect — Lepper et al., 1973, found children promised an award for drawing subsequently drew about half as much in free play), while unexpected rewards and verbal encouragement do not undermine. Points, badges, and leaderboards import exactly the risky category: they buy short-term engagement metrics and can crowd out interest in the learning itself — and engagement metrics are fluency-side measures, so by the performance/learning dissociation they are not evidence of durable learning in the first place. If game mechanics are used at all, they should reward learning behaviors (completing a hard retrieval streak) rather than performance outcomes, and celebration should be informational ("you retained 90% of last month's concepts") rather than controlling.

Interest develops in phases. Hidi & Renninger (2006): triggered situational interest (a hook — novelty, story, surprise) → maintained situational interest (sustained by meaning and personal relevance) → emerging individual interest (the student starts returning voluntarily) → well-developed individual interest (self-sustaining curiosity that survives difficulty). External triggers only start the process; the transition to individual interest requires connecting material to the learner's own projects and identity over long periods. This is the strongest research argument for tutors that persist across years rather than sessions.

Self-efficacy (Bandura, 1977) — task-specific belief in one's capability — predicts persistence and accounts for ~14% of variance in academic performance (r = .38; Multon, Brown & Lent, 1991). Its dominant source is mastery experiences, i.e., genuinely succeeding at genuinely hard things — which the success-rate controller manufactures. Cheap praise is not a substitute (see the feedback levels above).

Mindset and grit, with replication honesty. Dweck's growth-mindset interventions replicate at much smaller magnitudes than the popular narrative: Sisk et al. (2018) found the mindset-achievement correlation is r = .10 and intervention effects average d = 0.08 (d = 0.19 for academically at-risk students, d = 0.34 for low-SES); the rigorous national study (Yeager et al., 2019) found ~0.10 GPA points, concentrated in lower-achieving students and strongest where peer norms supported challenge-seeking. Grit (Duckworth et al., 2007) is largely conscientiousness re-measured (Credé et al., 2017: meta-analytic grit-conscientiousness correlation ρ = .84; grit-academic-performance ρ = .18). The defensible takeaway is narrow: feedback framing matters — attribute struggle to strategy and process, treat errors as information, never signal that ability is fixed — but do not expect mindset messaging to substitute for the cognitive interventions.

What Not to Do

The negative results are as load-bearing as the positive ones. Each row is a documented failure mode, not a style preference:

Two of these deserve emphasis because they are the ones AI tutors are currently getting wrong. Help abuse / gaming the system: Baker, Corbett, Koedinger & Wagner (2004) documented students systematically exploiting hint mechanics to extract answers without thinking; gaming frequency correlated with post-test scores as strongly as prior knowledge did, and no other off-task behavior came close (the help-abuse link goes back to Aleven & Koedinger, 2000). An LLM tutor with a friendly chat interface is a dramatically better answer-extraction machine than 2004-era hint buttons — and Bastani et al. (2025) is the controlled demonstration that this failure mode is live: same model, same students, and the guardrailed variant helped while the unrestricted variant actively hurt exam performance. Hint ladders need rate structure, and the system needs to detect answer-fishing patterns. Short-horizon optimization: almost every engagement-driven product objective (session completion, streaks, time-on-task, satisfaction ratings) is a fluency-side metric, and fluency-side metrics select against desirable difficulties. The only honest objective is delayed retention — measured by the system's own spaced retrievals.

Design Implications: The Core Loop of a Durable-Learning Tutor

Assembling the evidence into an architecture, the load-bearing components are a persistent per-student knowledge state, a retrieval scheduler, and a pedagogy selector — with the LLM as the interaction layer, not the system of record:

# Persistent state per student (survives sessions; horizon = years): # knowledge graph: concepts + prerequisite edges (from curriculum) # per concept: mastery belief (BKT-style L with decay), stability S, # last_retrieval, error patterns, effective explanation styles # learner profile: interests, self-efficacy signals, help-use patterns, # Elo-style ability rating per topic each session: due = concepts where predicted_recall(now, S, last_retrieval) < 0.90 frontier = unmastered concepts whose prerequisites are all mastered agenda = interleave(due_retrievals, frontier_instruction) # interleave across confusable topics; bias gaps long when unsure for item in agenda: mode = select_pedagogy(item): new concept, low prior -> (optional bounded struggle / pretest question first, if age- appropriate) then direct instruction + worked example partial mastery -> completion problems, guided practice mastered, retrieval due -> effortful retrieval, then feedback; keep in queue past first success (3 recalls + 3 spaced relearnings) fragile (correct, low conf) -> "explain your reasoning" probe repeated failure -> walk prerequisite graph backward response = elicit(item, mode) # LLM: dialogue, hint ladder diagnosis = analyze(response) # LLM: error/misconception ID give process-level feedback # never person-level; delayed # feedback is fine, even good update L (Bayesian) and S (scheduler fit); update Elo; log evidence if answer-fishing / gaming detected: tighten hint ladder, flag periodically: run transfer checks on untested sibling material # Pan & Rickard surface the model to the student # metacognition: "you're about # to forget X; here's your curve" report to teacher: struggling students, decayed load-bearing concepts

The claims embedded in this loop, each traceable to a result above: retrieval events outrank new content (testing effect); items stay in the queue after first success (Karpicke & Roediger 2008); the schedule follows per-concept stability, not the curriculum calendar (ridgeline), erring long under uncertainty (asymmetric penalty); pedagogy mode is a function of the per-concept knowledge state (expertise reversal / guidance fading), with question-before-explanation as the default (pretesting); advancement requires demonstrated mastery plus scheduled reactivation (mastery learning + forgetting); feedback is process-level and always attached to retrieval (Rowland's feedback moderator; Hattie & Timperley); transfer is verified, not assumed (Pan & Rickard); the model is interpretable enough to justify decisions and to be surfaced to the student as metacognitive training (Schraw, 1998; Kornell & Bjork 2007's 18% statistic is the gap to close); and the teacher stays in the loop as the relational and motivational anchor, with the system as force multiplier.

What this architecture does not solve — the honest open problems: the motivation cliff (spaced retrieval works and is disliked; narrative framing and visible retention wins are mitigations, not solutions), the cold start (a new student's parameters are unknown; diagnostic pretests, population priors, and a few Elo-calibration items only roughly help), the relationship problem (no deployed system maintains a coherent, trusted persona and student model over years — the gap between current tooling and a Diamond Age Primer is precisely persistence, not language capability), and measurement (distinguishing real mastery from fragile knowledge at scale still lacks a validated, cheap instrument; LLM-graded explanations are promising and unproven — and note that no LLM-tutor RCT has yet reported delayed-retention outcomes, so the field is still grading itself on the performance side of the performance/learning dissociation). These four, not explanation generation, are where the research frontier actually is.

Sources

• Adesope, O. O., Trevisan, D. A., & Sundararajan, N. (2017). Rethinking the use of tests: A meta-analysis of practice testing. Review of Educational Research, 87(3), 659–701.
• Agarwal, P. K., & Bain, P. M. (2019). Powerful Teaching: Unleash the Science of Learning. Jossey-Bass.
• Aleven, V., & Koedinger, K. R. (2000). Limitations of student control: Do students know when they need help? ITS 2000, 292–303.
• Baker, R. S., Corbett, A. T., Koedinger, K. R., & Wagner, A. Z. (2004). Off-task behavior in the Cognitive Tutor classroom: When students "game the system." CHI 2004, 383–390.
• Baker, R. S., Corbett, A. T., & Aleven, V. (2008). More accurate student modeling through contextual estimation of slip and guess probabilities in Bayesian knowledge tracing. ITS 2008.
• Bandura, A. (1977). Self-efficacy: Toward a unifying theory of behavioral change. Psychological Review, 84(2), 191–215.
• Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö., & Mariman, R. (2025). Generative AI can harm learning. PNAS, 122(26).
• Beck, J. E., & Chang, K. (2007). Identifiability: A fundamental problem of student modeling. User Modeling 2007, LNCS 4511, 137–146.
• Bergeron, P.-J., & Rivard, L. (2017). How to engage in pseudoscience with real data: A criticism of John Hattie's arguments in Visible Learning from the perspective of a statistician. McGill Journal of Education, 52(1).
• Birnbaum, M. S., Kornell, N., Bjork, E. L., & Bjork, R. A. (2013). Why interleaving enhances inductive learning: The roles of discrimination and retrieval. Memory & Cognition, 41(3), 392–402.
• Bisra, K., Liu, Q., Nesbit, J. C., Salimi, F., & Winne, P. H. (2018). Inducing self-explanation: A meta-analysis. Educational Psychology Review, 30(3), 703–725.
• Bjork, R. A. (1994). Memory and metamemory considerations in the training of human beings. In Metacognition: Knowing about Knowing (pp. 185–205). MIT Press.
• Bjork, R. A., & Bjork, E. L. (1992). A new theory of disuse and an old theory of stimulus fluctuation. In From Learning Processes to Cognitive Processes (Vol. 2, pp. 35–67). Erlbaum.
• Bloom, B. S. (1968). Learning for mastery. Evaluation Comment, 1(2), 1–12.
• Bloom, B. S. (1984). The 2 sigma problem: The search for methods of group instruction as effective as one-to-one tutoring. Educational Researcher, 13(6), 4–16.
• Brunmair, M., & Richter, T. (2019). Similarity matters: A meta-analysis of interleaved learning and its moderators. Psychological Bulletin, 145(11), 1029–1052.
• Butler, A. C., & Roediger, H. L. (2008). Feedback enhances the positive effects and reduces the negative effects of multiple-choice testing. Memory & Cognition, 36(3), 604–616.
• Carroll, J. B. (1963). A model of school learning. Teachers College Record, 64(8), 723–733.
• Carvalho, P. F., & Goldstone, R. L. (2014). Putting category learning in order: Category structure and temporal arrangement affect the benefit of interleaved over blocked study. Memory & Cognition, 42(3), 481–495.
• Cepeda, N. J., Pashler, H., Vul, E., Wixted, J. T., & Rohrer, D. (2006). Distributed practice in verbal recall tasks: A review and quantitative synthesis. Psychological Bulletin, 132(3), 354–380.
• Cepeda, N. J., Vul, E., Rohrer, D., Wixted, J. T., & Pashler, H. (2008). Spacing effects in learning: A temporal ridgeline of optimal retention. Psychological Science, 19(11), 1095–1102.
• Chi, M. T. H., & Wylie, R. (2014). The ICAP framework: Linking cognitive engagement to active learning outcomes. Educational Psychologist, 49(4), 219–243.
• Corbett, A. T., & Anderson, J. R. (1994). Knowledge tracing: Modeling the acquisition of procedural knowledge. User Modeling and User-Adapted Interaction, 4(4), 253–278.
• Cowan, N. (2001). The magical number 4 in short-term memory: A reconsideration of mental storage capacity. Behavioral and Brain Sciences, 24(1), 87–114.
• Credé, M., Tynan, M. C., & Harms, P. D. (2017). Much ado about grit: A meta-analytic synthesis of the grit literature. Journal of Personality and Social Psychology, 113(3), 492–511.
• De Simone, M., Tiberti, F., Barron Rodriguez, M., et al. (2025). From Chalkboards to Chatbots: Evaluating a GenAI Intervention in Nigeria. World Bank Policy Research Working Paper 11125.
• Deci, E. L., Koestner, R., & Ryan, R. M. (1999). A meta-analytic review of experiments examining the effects of extrinsic rewards on intrinsic motivation. Psychological Bulletin, 125(6), 627–668.
• Duckworth, A. L., Peterson, C., Matthews, M. D., & Kelly, D. R. (2007). Grit: Perseverance and passion for long-term goals. Journal of Personality and Social Psychology, 92(6), 1087–1101.
• Dunlosky, J., Rawson, K. A., Marsh, E. J., Nathan, M. J., & Willingham, D. T. (2013). Improving students' learning with effective learning techniques. Psychological Science in the Public Interest, 14(1), 4–58.
• Ebbinghaus, H. (1885). Über das Gedächtnis. Duncker & Humblot.
• Education Endowment Foundation. Teaching and Learning Toolkit: One-to-one tuition. educationendowmentfoundation.org.uk.
• Gervet, T., Koedinger, K., Schneider, J., & Mitchell, T. (2020). When is deep learning the best approach to knowledge tracing? Journal of Educational Data Mining, 12(3), 31–54.
• Ghosh, A., Heffernan, N., & Lan, A. S. (2020). Context-aware attentive knowledge tracing. KDD 2020.
• Guskey, T. R. (2010). Lessons of mastery learning. Educational Leadership, 68(2), 52–57.
• Hattie, J. (2009). Visible Learning. Routledge. (Updated influence lists: visible-learning.org.)
• Hattie, J., & Timperley, H. (2007). The power of feedback. Review of Educational Research, 77(1), 81–112.
• Hidi, S., & Renninger, K. A. (2006). The four-phase model of interest development. Educational Psychologist, 41(2), 111–127.
• Kalyuga, S., Ayres, P., Chandler, P., & Sweller, J. (2003). The expertise reversal effect. Educational Psychologist, 38(1), 23–31.
• Kang, S. H. K. (2016). Spaced repetition promotes efficient and effective learning. Policy Insights from the Behavioral and Brain Sciences, 3(1), 12–19.
• Kang, S. H. K., & Pashler, H. (2012). Learning painting styles: Spacing is advantageous when it promotes discriminative contrast. Applied Cognitive Psychology, 26(1), 97–103.
• Karpicke, J. D., & Blunt, J. R. (2011). Retrieval practice produces more learning than elaborative studying with concept mapping. Science, 331(6018), 772–775.
• Karpicke, J. D., & Roediger, H. L. (2007). Expanding retrieval practice promotes short-term retention, but equally spaced retrieval enhances long-term retention. Journal of Experimental Psychology: Learning, Memory, and Cognition, 33(4), 704–719.
• Karpicke, J. D., & Roediger, H. L. (2008). The critical importance of retrieval for learning. Science, 319(5865), 966–968.
• Kestin, G., Miller, K., Klales, A., Milbourne, T., & Ponti, G. (2025). AI tutoring outperforms in-class active learning. Scientific Reports, 15, 17458.
• Khajah, M., Lindsey, R. V., & Mozer, M. C. (2016). How deep is knowledge tracing? EDM 2016.
• Kirschner, P. A., Sweller, J., & Clark, R. E. (2006). Why minimal guidance during instruction does not work. Educational Psychologist, 41(2), 75–86.
• Klinkenberg, S., Straatemeier, M., & van der Maas, H. L. J. (2011). Computer adaptive practice of maths ability using a new item response model for on the fly ability and difficulty estimation. Computers & Education, 57(2), 1813–1824.
• Kluger, A. N., & DeNisi, A. (1996). The effects of feedback interventions on performance. Psychological Bulletin, 119(2), 254–284.
• Kornell, N., & Bjork, R. A. (2007). The promise and perils of self-regulated study. Psychonomic Bulletin & Review, 14(2), 219–224.
• Kornell, N., & Bjork, R. A. (2008). Learning concepts and categories: Is spacing the "enemy of induction"? Psychological Science, 19(6), 585–592.
• Kornell, N., Hays, M. J., & Bjork, R. A. (2009). Unsuccessful retrieval attempts enhance subsequent learning. Journal of Experimental Psychology: Learning, Memory, and Cognition, 35(4), 989–998.
• Kulik, C.-L. C., Kulik, J. A., & Bangert-Drowns, R. L. (1990). Effectiveness of mastery learning programs: A meta-analysis. Review of Educational Research, 60(2), 265–299.
• Kulik, C.-L. C., & Kulik, J. A. (1991). Effectiveness of computer-based instruction: An updated analysis. Computers in Human Behavior, 7(1-2), 75–94.
• Kulik, J. A., & Fletcher, J. D. (2016). Effectiveness of intelligent tutoring systems: A meta-analytic review. Review of Educational Research, 86(1), 42–78.
• Landauer, T. K., & Bjork, R. A. (1978). Optimum rehearsal patterns and name learning. In Practical Aspects of Memory (pp. 625–632). Academic Press.
• Lepper, M. R., Greene, D., & Nisbett, R. E. (1973). Undermining children's intrinsic interest with extrinsic reward. Journal of Personality and Social Psychology, 28(1), 129–137.
• Lindsey, R. V., Shroyer, J. D., Pashler, H., & Mozer, M. C. (2014). Improving students' long-term knowledge retention through personalized review. Psychological Science, 25(3), 639–647.
• Ma, W., Adesope, O. O., Nesbit, J. C., & Liu, Q. (2014). Intelligent tutoring systems and learning outcomes: A meta-analysis. Journal of Educational Psychology, 106(4), 901–918.
• Mazza, S., Gerbier, E., Gustin, M.-P., Kasikci, Z., Koenig, O., Toppino, T. C., & Magnin, M. (2016). Relearn faster and retain longer: Along with practice, sleep makes perfect. Psychological Science, 27(10), 1321–1330.
• McDaniel, M. A., Agarwal, P. K., Huelser, B. J., McDermott, K. B., & Roediger, H. L. (2011). Test-enhanced learning in a middle school science classroom. Journal of Educational Psychology, 103(2), 399–414.
• Multon, K. D., Brown, S. D., & Lent, R. W. (1991). Relation of self-efficacy beliefs to academic outcomes. Journal of Counseling Psychology, 38(1), 30–38.
• Murray, T., & Arroyo, I. (2002). Toward measuring and maintaining the zone of proximal development in adaptive instructional systems. ITS 2002, 749–758.
• Murre, J. M. J., & Dros, J. (2015). Replication and analysis of Ebbinghaus' forgetting curve. PLOS ONE, 10(7), e0120644.
• Nickow, A., Oreopoulos, P., & Quan, V. (2020). The impressive effects of tutoring on preK-12 learning: A systematic review and meta-analysis of the experimental evidence. NBER Working Paper 27476. (Published 2024 in American Educational Research Journal.)
• Pan, S. C., & Rickard, T. C. (2018). Transfer of test-enhanced learning: Meta-analytic review and synthesis. Psychological Bulletin, 144(7), 710–756.
• Pandey, S., & Karypis, G. (2019). A self-attentive model for knowledge tracing. EDM 2019.
• Pane, J. F., Griffin, B. A., McCaffrey, D. F., & Karam, R. (2014). Effectiveness of Cognitive Tutor Algebra I at scale. Educational Evaluation and Policy Analysis, 36(2), 127–144.
• Pashler, H., McDaniel, M., Rohrer, D., & Bjork, R. (2008). Learning styles: Concepts and evidence. Psychological Science in the Public Interest, 9(3), 105–119.
• Pavlik, P. I., Cen, H., & Koedinger, K. R. (2009). Performance factors analysis — a new alternative to knowledge tracing. AIED 2009.
• Pelánek, R. (2016). Applications of the Elo rating system in adaptive educational systems. Computers & Education, 98, 169–179.
• Piech, C., Bassen, J., Huang, J., Ganguli, S., Sahami, M., Guibas, L. J., & Sohl-Dickstein, J. (2015). Deep knowledge tracing. NeurIPS 2015.
• Pyc, M. A., & Rawson, K. A. (2009). Testing the retrieval effort hypothesis. Journal of Memory and Language, 60(4), 437–447.
• Rawson, K. A., & Dunlosky, J. (2011). Optimizing schedules of retrieval practice for durable and efficient learning: How much is enough? Journal of Experimental Psychology: General, 140(3), 283–302.
• Renkl, A., & Atkinson, R. K. (2003). Structuring the transition from example study to problem solving in cognitive skill acquisition. Educational Psychologist, 38(1), 15–22.
• Richland, L. E., Kornell, N., & Kao, L. S. (2009). The pretesting effect: Do unsuccessful retrieval attempts enhance learning? Journal of Experimental Psychology: Applied, 15(3), 243–257.
• Roediger, H. L., & Karpicke, J. D. (2006). Test-enhanced learning: Taking memory tests improves long-term retention. Psychological Science, 17(3), 249–255.
• Roediger, H. L., Agarwal, P. K., McDaniel, M. A., & McDermott, K. B. (2011). Test-enhanced learning in the classroom: Long-term improvements from quizzing. Journal of Experimental Psychology: Applied, 17(4), 382–395.
• Rohrer, D. (2012). Interleaving helps students distinguish among similar concepts. Educational Psychology Review, 24, 355–367.
• Rohrer, D., & Taylor, K. (2007). The shuffling of mathematics problems improves learning. Instructional Science, 35(6), 481–498.
• Rohrer, D., Dedrick, R. F., & Stershic, S. (2015). Interleaved practice improves mathematics learning. Journal of Educational Psychology, 107(3), 900–908.
• Rohrer, D., Dedrick, R. F., Hartwig, M. K., & Cheung, C.-N. (2020). A randomized controlled trial of interleaved mathematics practice. Journal of Educational Psychology, 112(1), 40–52.
• Rowland, C. A. (2014). The effect of testing versus restudy on retention: A meta-analytic review of the testing effect. Psychological Bulletin, 140(6), 1432–1463.
• Schraw, G. (1998). Promoting general metacognitive awareness. Instructional Science, 26(1–2), 113–125.
• Seabrook, R., Brown, G. D., & Solity, J. E. (2005). Distributed and massed practice: From laboratory to classroom. Applied Cognitive Psychology, 19(1), 107–122.
• Settles, B., & Meeder, B. (2016). A trainable spaced repetition model for language learning. ACL 2016, 1848–1858.
• Sinha, T., & Kapur, M. (2021). When problem solving followed by instruction works: Evidence for productive failure. Review of Educational Research, 91(5), 761–798.
• Sisk, V. F., Burgoyne, A. P., Sun, J., Butler, J. L., & Macnamara, B. N. (2018). To what extent and under which circumstances are growth mind-sets important to academic achievement? Psychological Science, 29(4), 549–571.
• Slamecka, N. J., & Graf, P. (1978). The generation effect: Delineation of a phenomenon. Journal of Experimental Psychology: Human Learning and Memory, 4(6), 592–604.
• Slavin, R. E. (1987). Mastery learning reconsidered. Review of Educational Research, 57(2), 175–213.
• Smith, M. A., & Karpicke, J. D. (2014). Retrieval practice with short-answer, multiple-choice, and hybrid tests. Memory, 22(7), 784–802.
• Soderstrom, N. C., & Bjork, R. A. (2015). Learning versus performance: An integrative review. Perspectives on Psychological Science, 10(2), 176–199.
• Sweller, J. (1988). Cognitive load during problem solving: Effects on learning. Cognitive Science, 12(2), 257–285.
• Sweller, J., & Cooper, G. A. (1985). The use of worked examples as a substitute for problem solving in learning algebra. Cognition and Instruction, 2(1), 59–89.
• Sweller, J., van Merriënboer, J. J. G., & Paas, F. (1998; updated 2019). Cognitive architecture and instructional design. Educational Psychology Review, 10(3), 251–296; 31, 261–292.
• Tabibian, B., Upadhyay, U., De, A., Zarezade, A., Schölkopf, B., & Gomez-Rodriguez, M. (2019). Enhancing human learning via spaced repetition optimization. PNAS, 116(10), 3988–3993.
• VanLehn, K. (2011). The relative effectiveness of human tutoring, intelligent tutoring systems, and other tutoring systems. Educational Psychologist, 46(4), 197–221.
• Vansteenkiste, M., Simons, J., Lens, W., Sheldon, K. M., & Deci, E. L. (2004). Motivating learning, performance, and persistence. Journal of Personality and Social Psychology, 87(2), 246–260.
• Vygotsky, L. S. (1978). Mind in Society. Harvard University Press.
• Wilson, K. H., Karklin, Y., Han, B., & Ekanadham, C. (2016). Back to the basics: Bayesian extensions of IRT outperform neural networks for proficiency estimation. EDM 2016, 539–544.
• Wisniewski, B., Zierer, K., & Hattie, J. (2020). The power of feedback revisited: A meta-analysis of educational feedback research. Frontiers in Psychology, 10, 3087.
• Wozniak, P. (1987/1990). The SM-2 algorithm. SuperMemo; super-memory.com/english/ol/sm2.htm.
• Xiong, X., Zhao, S., Van Inwegen, E. G., & Beck, J. E. (2016). Going deeper with deep knowledge tracing. EDM 2016, 545–550.
• Yeager, D. S., et al. (2019). A national experiment reveals where a growth mindset improves achievement. Nature, 573, 364–369.
• Ye, J., Su, J., & Cao, Y. (2022). A stochastic shortest path algorithm for optimizing spaced repetition scheduling. KDD 2022, 4381–4390. (DSR model underlying FSRS; benchmark: github.com/open-spaced-repetition/srs-benchmark.)
• Zhang, J., Shi, X., King, I., & Yeung, D.-Y. (2017). Dynamic key-value memory networks for knowledge tracing. WWW 2017.

Research provenance: distilled from the deep-research reports in brain/inbox/dr/2026-07-24-long-term-learning/ (synthesis.md, comprehensive-report.md). All quantitative claims were subsequently verified against the primary papers (full texts or publisher abstracts) in a six-stream verification pass on 2026-08-02; numbers in the generated reports that disagreed with the original papers were corrected here, including the Rohrer interleaving scores, the VanLehn substep effect size (0.40, not 0.95), the Kulik mastery-learning figures, and the Cepeda gap-ratio guidance.