Techniques that work, checked for high-ability JEE students

For strong Grade 11-12 students preparing for a speeded, negatively-marked MCQ exam, four things have the best evidence and the most direct fit: practice testing in exam format with corrective feedback, review spaced against the actual exam date, mixed (interleaved) practice that forces the student to identify the problem type, and an explicit attempt rule for negative marking (under JEE Main’s +4/-1 scoring, answer any MCQ where the chance of being right is above 20%). Two techniques can backfire for students who already know the material. Worked examples help novices but lose their edge, or reverse, as expertise grows (Kalyuga 2007). Interleaving helped low scorers but showed signs of harm at the top in the only full-year RCT in a low-income setting (van der Haar et al. 2023). The Learning Framework is right on direction but overstates several magnitudes: its Muller (2008) citation points to the wrong paper, its Cepeda (2008) spacing rule for an 18-month horizon is an extrapolation the paper does not make, its Mindspark benchmark comes from below-grade-level middle-school students, and mastery learning’s “1 sigma” rests on three-week studies with narrow tests. The coach should treat each technique as a hypothesis to measure on Avanti’s own mock data, not a guaranteed effect.

Scope note. Every row below was opened and read in a downloaded copy on 2026-10-26. Most of this literature is from US undergraduates or US middle schools. Very little is from Indian adolescents, and almost none is from students preparing for a high-stakes competitive exam. Several works the task named could not be obtained and appear only under “Leads, not opened” at the end; their findings are not used as fact here.

Evidence table

Study	Design and population	n	Effect (unit)	What it means for the coach
Dunlosky et al. 2013	Narrative review of 10 study techniques, mostly lab and undergraduate studies	Hundreds of studies	Utility ratings: practice testing and distributed practice “high”; interleaved practice, self-explanation, elaborative interrogation “moderate”; rereading, highlighting, summarisation “low”	Default student time to testing and spacing, not rereading notes or rereading explanations
Roediger and Karpicke 2006	Lab experiments, US undergraduates, prose passages, recall tests without feedback	120 in Exp. 1	After 1 week: tested 56% vs restudied 42% recalled (d = 0.83). After 5 minutes restudy won, 81% vs 75%	Testing looks worse in the short run and better a week later. Do not judge a practice session by same-day accuracy
Rowland 2014	Meta-analysis of testing vs restudy, mostly lab	159 effect sizes	g = 0.50 overall; recognition (MCQ-type) initial tests g = 0.29 vs cued recall g = 0.61; retention interval of 1 day or more g = 0.69	Recall-style prompts may beat MCQ prompts for learning, even though the exam is MCQ
Adesope et al. 2017	Meta-analysis of practice tests vs other conditions	272 independent effect sizes	g = 0.61 overall; g = 0.51 vs restudy; secondary students g = 0.83; classroom g = 0.67; with feedback g = 0.63, without g = 0.60; MCQ practice g = 0.70, short answer g = 0.48	Practice testing works across levels and settings. It conflicts with Rowland on MCQ vs recall, so test both formats
Agarwal et al. 2012	Review of classroom experiments, US middle-school social studies, within-student quizzed vs not quizzed material	Several classes	End of semester: 79% vs 67% correct on quizzed vs non-quizzed material	Low-stakes quizzing in a real school transfers to later unit exams
van Gog and Sweller 2015	Review of a journal special issue on testing with complex materials	Several experiments	Four experiments on learning circuit-fault diagnosis and probability problems from worked examples: no testing effect in any	For multi-step JEE problem solving, the testing effect is not guaranteed. Mix solving with studying solutions
Cepeda et al. 2008	Internet experiment, adults learning obscure facts, one review session, gap up to 3.5 months	Over 1,350	Optimal gap was about 20% of the test delay for delays of a few weeks, falling to about 5% at one year; for a 350-day delay the best gap tested was 21 days (recall)	Schedule reviews against the real exam date; the right gap shrinks as a fraction of the horizon
Bego et al. 2024	Classroom experiment, 9 US introductory STEM courses, spaced vs massed quiz items within students	578	Significant spacing benefit in 2 of 9 courses; pooled effect significant only when calculus was included	Spacing in real STEM courses is fragile; informative feedback and recall-style items may be what made it work in calculus
Rohrer et al. 2015	Classroom experiment, US 7th-grade maths, interleaved vs blocked practice over 3 months	126	d = 0.42 at 1 day; d = 0.79 at 30 days	Mixed practice helps students choose the right method, the skill a mixed exam paper tests
Rohrer et al. 2020	Cluster RCT, 54 US 7th-grade maths classes, 4 months	787	Unannounced test one month later: 61% vs 38% (d = 0.83)	The strongest interleaving RCT; same age and subject limits as above
van der Haar et al. 2023	Cluster RCT, 62 Grade 5 classrooms, low-cost private schools, Lagos	62 classrooms	Short-term retention +0.29 SD; end-of-year cumulative test +0.04 SD (not significant); gains at the bottom of the distribution offset by negative effects at the top	Interleaving may not help, and may hurt, the strongest students on long-run cumulative tests. Measure by ability band
Kalyuga 2007	Review of expertise-reversal experiments (school and adult learners)	Many studies	Worked-example advantage disappeared with experience; in high-school coordinate geometry, less knowledgeable students gained more from worked examples and more knowledgeable students from problem solving	Give worked or faded examples on chapters the student does not know; give problems first on chapters they know
Wisniewski et al. 2020	Meta-analysis of feedback on learning	435 studies, N over 61,000	d = 0.48 overall; high-information feedback d = 0.99; corrective d = 0.46; reinforcement or punishment d = 0.24	What the feedback says matters more than whether it arrives. Explain why the error happened and what to do next
Schroeder and Kucera 2022	Meta-analysis of refutation text vs other texts	44 comparisons, n = 3,869	g = 0.41; trim-and-fill adjusted g = 0.28; dissertations g = 0.11 (not significant); 17 of 44 comparisons tested the same day (most delayed test coded), and timing did not moderate the effect	Naming and refuting a misconception helps, but by roughly a third of a standard deviation, not double
Sadler and Sonnert 2016	Observational, US middle-school physical science, teachers and students took the same MCQ items	9,556 students, 181 teachers	On items with a popular wrong answer, teachers who could predict that wrong answer had larger classroom gains than teachers who knew only the right answer	The popular distractor is diagnostic. The coach should name which wrong option a student chose and why it is tempting
Bloom 1984	Essay summarising two PhD experiments	Small samples	Tutored students about 2 SD above conventional class; mastery learning about 1 SD	The origin of the 2-sigma claim. See the next two rows before quoting it
von Hippel 2024	Re-reading of the dissertations behind Bloom 1984	Grades 4, 5, 8; 3-week experiments	2-sigma effects were on narrow tests of unfamiliar topics; about half the effect came from extra testing, feedback and retesting	Mastery and tutoring gains on broad exams will be far smaller. Keep the retest-after-feedback mechanism
Slavin 1987	Best-evidence synthesis of group-based mastery learning, studies of 4 weeks or more	Many studies	“Essentially no evidence” of effects on standardised tests; moderate positive effects on experimenter-made tests	A mastery gate may raise scores on its own items without moving mock scores. Judge it on mocks
Nickow et al. 2020	Meta-analysis of tutoring RCTs, preschool to secondary	96 studies	Pooled 0.37 SD	A realistic ceiling for intensive human tutoring, not for a weekly chat
Kraft 2020	Methods paper on interpreting education effect sizes	Review of RCTs	For standardised achievement outcomes: under 0.05 small, 0.05 to under 0.20 medium, 0.20 or more large	A pilot that moves mock scores by 0.10 SD would be a real result
Muralidharan et al. 2019 (working-paper copy)	RCT, after-school Mindspark centres, Delhi, grades 6-9, students far below grade level	619	ITT after 4.5 months: +0.36 SD maths, +0.22 SD Hindi	Not a benchmark for high-ability Grade 11-12 students; different population and baseline
Akyol, Key and Krishna 2016	Structural model on Turkish university entrance exam data (+1 / -0.25, five options)	National exam data	Women and high-ability students are more risk averse; removing negative marking would raise women’s share of the top 5% from 39.9% to 40.2%	Risk aversion makes strong students skip answerable questions
Baldiga 2014	Lab experiment, US, SAT II subject-test questions	406	With a small penalty women answered significantly fewer questions; no difference in knowledge or confidence; skippers scored lower conditional on knowledge	Coach girls in particular on when to attempt
Pekkarinen 2014 (IZA DP 8018)	Observational, Finnish economics and business entrance exams with negative marking	Applicant data	Women omitted more items and deviated further from the entry-maximising number of answers, by answering too few	Same pattern in a real high-stakes exam
Krishna, Akyol and Ozer 2025	Lab-in-field experiment, Grade 12 students, one Turkish high school, time limits varied	One school	Extra time helped men more; risk aversion did not differ significantly by gender	The gender-and-guessing finding is contested; measure it in Avanti data before targeting by gender
NTA 2026 JEE Main bulletin (for 2027) and a 2026 question paper	Official documents	n/a	Paper 1: 75 questions, 180 minutes, +4 correct, -1 incorrect, 0 unanswered, for both MCQ and numerical-value questions; MCQs carry four options	Break-even for answering an MCQ is a 20% chance of being right; blind guessing on a numerical question loses marks

What the evidence says

Retrieval practice (testing effect)

Taking a test on material improves later recall more than restudying it. In Roediger and Karpicke (2006), students who were tested recalled 56% of a prose passage after a week, against 42% for students who restudied (d = 0.83). The same students did worse at 5 minutes (75% vs 81%). That reversal matters for a coach: a session of retrieval will feel and look less productive on the day.

The meta-analyses agree on direction and differ on size. Rowland (2014) found g = 0.50 against restudy. Adesope et al. (2017) found g = 0.61 against all comparison conditions and g = 0.51 against restudy, with secondary students at g = 0.83 and classroom studies at g = 0.67. Agarwal et al. (2012) report a classroom case: end-of-semester performance of 79% on quizzed material against 67% on non-quizzed material in US middle-school social studies. Dunlosky et al. (2013) rate practice testing “high utility”.

The two meta-analyses disagree on format. Rowland (2014) found recognition (MCQ-type) initial tests weaker (g = 0.29) than cued recall (g = 0.61). Adesope et al. (2017) found the opposite (MCQ g = 0.70, short answer g = 0.48). Neither is about JEE.

The boundary that matters most for JEE is complexity. Van Gog and Sweller (2015) summarise a special issue in which four experiments on learning to solve circuit and probability problems from worked examples found no testing effect, and they argue the effect “decreases or even disappears” as the number of interacting elements rises. JEE problems are high in element interactivity. The safe reading is that retrieval works well for facts, definitions, formulas and concept checks, and is unproven as a substitute for studying solutions on multi-step problems.

Spacing

Cepeda et al. (2008) ran over 1,350 adults through fact learning with one review at gaps up to 3.5 months and final tests up to 350 days later. The best gap was about 20% of the test delay for short horizons, falling to about 5% at one year; for the 350-day delay the best gap tested was 21 days for recall. These are trivia facts with one review, so they set the shape of the schedule rather than its exact numbers for physics problem solving.

Classroom STEM evidence is weaker than lab evidence. Bego et al. (2024) embedded spaced vs massed retrieval in nine introductory STEM courses (578 students). Spacing helped significantly in two courses, and the pooled effect was significant only when calculus was included. They note the calculus course used fill-in-the-blank items with informative feedback, while most others used MCQ items with minimal feedback that students were not required to read. That is a design lesson for the coach, not proof of a mechanism.

Interleaving

Interleaving means mixing problem types so that consecutive problems need different methods. In two US 7th-grade maths studies it produced large effects on delayed tests: d = 0.79 at 30 days in 126 students (Rohrer et al. 2015) and 61% vs 38% (d = 0.83) one month later in a cluster RCT of 787 students in 54 classes (Rohrer et al. 2020). Rohrer et al. (2020) state that interleaving also guarantees spacing, so the two effects are confounded by design.

The only full-year RCT outside the US complicates this. Van der Haar et al. (2023) randomised 62 Grade 5 classrooms in Lagos. Interleaving raised short-term retention by 0.29 SD but had no detectable effect on an end-of-year cumulative test (0.04 SD), and quantile estimates suggested gains at the bottom were offset by negative effects at the top. The authors suggest, and label as speculative, that top students may already classify problems well and gain more from blocked practice that builds procedural accuracy. Dunlosky et al. (2013) rated interleaving only “moderate utility”.

Worked examples, problem solving and expertise reversal

Kalyuga (2007) reviews a body of experiments in which a technique that helps novices stops helping, or hurts, more knowledgeable learners. In Kalyuga et al. (2001, as summarised there) the advantage of worked examples over problem solving disappeared as trainees gained experience, and for complex tasks problem solving became better. In Kalyuga and Sweller (2004, as summarised there), less knowledgeable high-school students learned coordinate geometry better from worked examples and more knowledgeable students learned better from problem solving. In Reisslein’s engineering studies (as summarised there), knowledgeable learners did better with fast fading or immediate problem solving, and novices with slow fading. Kalyuga (2007) also describes “rapid diagnostic” tests, where a learner is shown a problem and asked for the first step, that correlated with longer tests (up to .92 in one study) and can route learners quickly.

For JNV CoE students this means the right format is chapter-specific. The same student can be a novice in rotational dynamics and an expert in kinematics.

Feedback

Wisniewski et al. (2020) pooled 435 studies (N over 61,000) and found d = 0.48, with very large differences by content: high-information feedback (what was wrong, why, and what to do next) d = 0.99; corrective feedback d = 0.46; reinforcement or punishment d = 0.24. Their estimates inherit the heterogeneity and quality issues of the Hattie programme they extend, so the ranking is more trustworthy than the absolute numbers. Adesope et al. (2017) found practice tests worked about equally with feedback (g = 0.63) and without (g = 0.60). On timing (immediate vs delayed), I did not obtain a primary study; see Leads.

Mastery learning and the 2-sigma claim

Bloom (1984) reported that tutored students scored about 2 SD above a conventional class and mastery-learning students about 1 SD above. Von Hippel (2024) went back to the two dissertations behind those numbers. They were three-week experiments with grade 4, 5 and 8 students, on topics chosen because they were unfamiliar, with tests tied to what was taught. Tutored students also got extra testing, feedback and a second quiz after scoring below 80% or 90%, and he estimates that this extra testing and feedback explains about half of the 2-sigma effect. Slavin (1987) found “essentially no evidence” that group-based mastery learning raised standardised test scores in studies of four weeks or more, with moderate positive effects on experimenter-made tests. Nickow et al. (2020) put the pooled effect of tutoring RCTs at 0.37 SD across 96 studies, with none near 2 SD.

The useful mechanism that survives is test, give targeted feedback, retest on new items. The claim that does not survive is that a mastery gate will produce effects of 1 SD on broad exams.

Refutation and misconception-targeting

Schroeder and Kucera (2022) pooled 44 comparisons (n = 3,869): refutation text beat other texts by g = 0.41, or g = 0.28 after trim-and-fill adjustment. Effects were largest in conference proceedings (g = 0.73), smaller in journal articles (g = 0.45) and not significant in dissertations (g = 0.11), which is a common sign of selective publication. The authors coded each study’s most delayed test: 17 of the 44 comparisons tested on the same day, 11 at two days to a week, 13 at eight days to a month, 2 beyond a month and 1 did not report it. The effect did not differ by timing (Q = 0.48, p = 0.98). The authors’ text says most studies used a same-day test, which their own moderator table does not bear out.

Sadler and Sonnert (2016), summarising their study of 9,556 students and 181 US middle-school science teachers, found that on items with a popular wrong answer, teachers who could predict that wrong answer produced larger classroom gains than teachers who only knew the right answer. This is observational. It supports using the specific distractor chosen, in addition to right or wrong, as the diagnostic signal.

Test-taking under negative marking

The JEE Main rules for 2027 are +4 for a correct answer, -1 for an incorrect one, 0 for unanswered, applied to both the MCQ section and the numerical-value section (NTA bulletin); the MCQs carry four options (2026 question paper). The expected marks from answering are 4p - (1 - p) = 5p - 1, which is positive whenever the chance p of being right exceeds 20%. A blind guess on a four-option MCQ (p = 0.25) is worth +0.25 marks in expectation, eliminating one option raises it to about +0.67, and eliminating two to +1.5. A blind guess on a numerical-value question has p close to zero and costs about 1 mark. This is arithmetic on the official scheme, not a research finding.

Several studies find that some groups skip too often. In Turkish entrance-exam data, Akyol, Key and Krishna (2016) estimate that women and high-ability students are more risk averse; in their model, removing the penalty would raise women’s share of the top 5% from 39.9% to 40.2%. That exam’s penalty (-0.25 with five options) makes a blind guess worth zero; JEE’s makes it worth +0.25, so skipping costs more in JEE. Baldiga (2014) found in a US lab experiment with SAT II questions that a small penalty made women answer significantly fewer questions, with no difference in knowledge or confidence, and that skippers scored lower conditional on knowledge. Pekkarinen (2014 working paper) found Finnish women applicants omitted more items and fell further short of the entry-maximising number of answers. Against this, Krishna, Akyol and Ozer (2025), in a lab-in-field experiment with Grade 12 students at one Turkish school, found no significant gender difference in risk aversion and found that extra time helped men more.

Calibrating expected effects

Kraft (2020) proposes that, for causal studies on standardised achievement, effects under 0.05 SD are small, 0.05 to under 0.20 medium, and 0.20 or more large. The Mindspark RCT (Muralidharan et al. 2019; the archived working-paper copy reports +0.36 SD maths and +0.22 SD Hindi ITT after 4.5 months) is strong Indian evidence for adaptive practice, but its students were in grades 6-9 and the average grade 6 student was about 2.5 grade levels behind. It does not tell us what a weekly AI coach will do for high-ability Grade 12 students in a residential CoE.

Where the evidence is weak or contested

1. Population mismatch. Almost every primary study here is on US undergraduates, US 7th graders or lab participants. None is on Indian Grade 11-12 students preparing for JEE. Effect sizes should be read as direction, not magnitude.
2. Complex problem solving. The testing effect is well established for facts and prose. For multi-step problem solving it is contested (van Gog and Sweller 2015). JEE is mostly the second kind.
3. Interleaving for top students. Large US effects (Rohrer et al. 2015, 2020) vs a null long-run average and possible harm at the top in Lagos (van der Haar et al. 2023). The mechanism offered for the top-student result is the authors’ speculation.
4. MCQ vs recall as a practice format. Rowland (2014) and Adesope et al. (2017) point in opposite directions.
5. Spacing in real STEM classrooms. Bego et al. (2024) found it in 2 of 9 courses.
6. Mastery learning. Large effects come from short studies with narrow tests (Bloom 1984, as re-read by von Hippel 2024); broad standardised measures show little (Slavin 1987). I could not obtain Kulik et al. (1990), the main pro-mastery meta-analysis, so its estimate is not used.
7. Refutation magnitude. g = 0.41, 0.28 after bias adjustment, and near zero in unpublished dissertations (Schroeder and Kucera 2022).
8. Gender and guessing. Three studies find women omit more; the most recent finds no gender difference in risk aversion (Krishna, Akyol and Ozer 2025). None is from India.
9. Feedback timing, productive failure and peer instruction. I could not obtain the primary sources (see Leads), so I make no claims about them.

Design rules for the Avanti coach

1. Make retrieval the default session activity for concepts, formulas and concept checks. Ask the student to recall or solve before showing anything. Expect same-day accuracy to look worse than rereading (Roediger and Karpicke 2006; Adesope et al. 2017).
2. Do not replace solution study with testing on multi-step problems. For a new problem type, alternate a worked solution with a similar problem to solve, rather than testing cold (van Gog and Sweller 2015; Kalyuga 2007).
3. Set the format per chapter level, not per student. On chapters where the student is at L0-L1 (handoff levels), start with a worked or faded example and then a near-identical problem. On chapters at L3-L4, give the problem first and offer the worked solution only after an attempt (Kalyuga 2007). Use a one-problem “what is your first step?” probe to decide (Kalyuga 2007).
4. Feedback must say why and what next. After a wrong answer, name the distractor chosen, the reasoning that makes it tempting, where that reasoning breaks, and one follow-up item. Then retest on a new item (Wisniewski et al. 2020; Sadler and Sonnert 2016; von Hippel 2024 on retest-after-feedback).
5. Use refutation, but budget for a modest effect. When a chapter shows a common wrong option across students, run a short refutation (state the misconception, why it feels right, where it fails), and check it a week later rather than the same day (Schroeder and Kucera 2022).
6. Schedule review from the exam date backwards. For each chapter, compute the remaining time to JEE and space reviews so that gaps shrink as a fraction of that horizon: weeks apart when the exam is months away, days apart in the final weeks (Cepeda et al. 2008). Keep a re-test of old chapters in every weekly plan (Bego et al. 2024).
7. Mix problem types in practice sets once each type has been learned once. Interleave across chapters in timed sets, as a mixed JEE paper does (Rohrer et al. 2015, 2020). Report interleaving effects separately for the top third of the class, and allow blocked consolidation for a top student whose accuracy on a known type is dropping (van der Haar et al. 2023).
8. Teach and track an explicit attempt rule. For JEE Main MCQs: attempt when you can eliminate at least one option or are more than 20% sure; never enter a blind guess on a numerical-value question (NTA bulletin arithmetic). Track, per student, the number of skipped MCQs where they had eliminated options, and flag students who skip far more than the rule implies (Akyol et al. 2016; Baldiga 2014; Pekkarinen 2014). Check whether the pattern differs by gender in Avanti mocks before acting on it, since the evidence is mixed (Krishna, Akyol and Ozer 2025).
9. Do not promise large effects. Set pilot targets in the Kraft (2020) range: 0.05 to 0.20 SD on mock scores is a meaningful result. Do not cite 2 SD (Bloom 1984; von Hippel 2024) or Mindspark’s 0.36 SD as a target for this population (Muralidharan et al. 2019).
10. Judge a mastery gate on mock scores, not on its own items. A gate can raise scores on the items it trains without moving broad exams (Slavin 1987). If the gate is kept, retest with new items after feedback, which is the part of the Bloom design that appears to carry the effect (von Hippel 2024).

Where the current spec conflicts with the evidence

1. Spec section 3, “bias towards retrieval and practice over explanation”. Right for concepts and for chapters the student knows. For a chapter a student is new to, the evidence favours worked or faded examples first (Kalyuga 2007), and testing alone on complex problems may not help (van Gog and Sweller 2015). The spec’s “worked answers are given, but at the right moment” should be made concrete: the moment is set by the student’s level on that chapter, not by the session type.
2. Spec section 8c, “Claude’s quizzes follow the same recency tiers” (50/30/20 weekly, 20/40/40 monthly). These splits are design choices; no opened source derives them. The research gives a rule for gaps relative to the time left before the exam (Cepeda et al. 2008). The coach should schedule by days-to-JEE per chapter.
3. Spec section 8c, mastery gate “randomised 8-10 ConcepTests, 80%+”. With 10 items, a student whose true accuracy is 80% passes only 68% of the time, and a student at 70% passes 38% of the time (binomial arithmetic; with 8 items and a 7/8 bar the figures are 50% and 26%). A short gate will misroute many students. The evidence also does not show that gating moves broad exam scores (Slavin 1987). Use the gate as a routing signal with retests, not as a hard block.
4. Spec section 8c, “calibrated expectations (Indian evidence 0.15-0.37 SD; 0.15-0.20 SD vs matched control is a realistic pilot target)”. The 0.37 comes from below-grade-level middle-school students in after-school centres (Muralidharan et al. 2019). For a 10-week, one-centre pilot with a matched (not randomised) comparison, 0.15-0.20 SD would be “large” on Kraft’s (2020) scale, and a matched design cannot support causal language.
5. Handoff concept, levels L1-L4 set by accuracy alone (40%, 70%). Under +4/-1 scoring, a student’s score depends on attempt rate as well as accuracy. A student with 75% accuracy who skips half the paper can score below one with 60% accuracy who attempts everything sensible. Levels should record attempt rate and skipped-but-answerable items alongside accuracy (NTA bulletin arithmetic; Akyol et al. 2016). The same binomial noise as point 3 applies to small per-chapter item counts.
6. Spec section 8, the AI Academic Report’s “negative-marking logic”. It should state the 20% break-even for MCQs and the no-blind-guess rule for numerical questions, which both carry -1 from the current bulletin. Any older advice that numerical questions are penalty-free is out of date (NTA bulletin).
7. Spec section 4.2, “problem-type mastery” from attempt history. Consistent with the evidence, provided practice sets mix types after first learning (Rohrer et al. 2020) and results are checked separately for the strongest students (van der Haar et al. 2023).

Learning Framework claims that the opened sources do not support as written

• Muller (2008), “learning gains approximately twice as large”, d = 0.83 and 0.79. The framework cites the paper as JCAL 24(5), 408-422 with DOI 10.1111/j.1365-2729.2008.00279.x. Crossref metadata for that DOI is a different paper (Westera et al., “Serious games for higher education”). The correct record is Muller, Bewes, Sharma and Reimann, JCAL 24(2), 144-155 (2008), DOI 10.1111/j.1365-2729.2007.00248.x. I could not obtain the paper or the thesis, so the d values are unverified. The meta-analytic average for refutation text is g = 0.41, or 0.28 after bias adjustment (Schroeder and Kucera 2022). “Twice as large” should not be repeated.
• Cepeda et al. (2008), “for retention intervals of approximately 18 months, the optimal review gap is 6-8 weeks”. The paper’s longest retention interval was 350 days, where the best gap tested was 21 days and the stated ratio was about 5%. It did not test 18 months, and its materials were trivia facts with a single review. Extrapolating its declining ratio would give a gap shorter than 6-8 weeks.
• Mindspark as “Indian secondary students”. The sample was grades 6-9 (middle school) and far below grade level (Muralidharan et al. 2019). The archived working-paper copy reports 0.36 and 0.22 SD; the framework’s 0.37 and 0.23 may come from the published version, which I did not open.
• “Over 100 years of cognitive science confirms” retrieval beats rereading. True for facts and prose (Rowland 2014; Adesope et al. 2017). Not established for complex problem solving (van Gog and Sweller 2015).
• “Unpredicted retrieval produces stronger retention than predicted retrieval.” No source is cited and I found none among the opened works. Treat as unsupported.
• Mastery gating and Bloom. The framework does not quote 2 sigma, but its mastery gate inherits Bloom’s design. The large effects come from three-week studies with narrow tests (von Hippel 2024) and do not appear on standardised tests (Slavin 1987).
• TBL: “immediate feedback produces stronger learning than delayed feedback.” Contested in the literature; I could not obtain the primary meta-analysis (Kulik and Kulik 1988). Do not state it as settled.
• Peer instruction “FCI gains approximately twice as large”. Not verified; Crouch and Mazur (2001) could not be obtained.

References

All opened 2026-10-26. Archived files are in scratchpad/sources/learning_science/ for this session (to be published to gs://avantifellows-research-sources/ai-coach/).

• Adesope, O. O., Trevisan, D. A., and Sundararajan, N. (2017). Rethinking the use of tests: A meta-analysis of practice testing. Review of Educational Research, 87(3), 659-701. https://doi.org/10.3102/0034654316689306. Opened 2026-10-26. File: adesope2017_meta.pdf (from Avanti archive nvs-lakshya/).
• Agarwal, P. K., Bain, P. M., and Chamberlain, R. W. (2012). The value of applied research: Retrieval practice improves classroom learning and recommendations from a teacher, a principal, and a scientist. Educational Psychology Review, 24, 437-448. https://doi.org/10.1007/s10648-012-9210-2. Opened 2026-10-26. File: agarwal2012.pdf (Avanti archive).
• Akyol, S. P., Key, J., and Krishna, K. (2016). Hit or miss? Test taking behavior in multiple choice exams. NBER Working Paper 22401. https://www.nber.org/papers/w22401. Opened 2026-10-26. File: akyol2016_hitormiss.pdf (Avanti archive).
• Baldiga, K. (2014). Gender differences in willingness to guess. Management Science (working-paper copy archived by Avanti). https://doi.org/10.1287/mnsc.2013.1776. Opened 2026-10-26. File: baldiga2014_guess.pdf (Avanti archive).
• Bego, C. R., Lyle, K. B., et al. (2024). Single-paper meta-analyses of the effects of spaced retrieval practice in nine introductory STEM courses: is the glass half full or half empty? International Journal of STEM Education, 11. https://doi.org/10.1186/s40594-024-00468-5. Opened 2026-10-26 in browser (PDF download blocked by a client challenge). File: bego2024_spaced_retrieval_stem_excerpt.txt (verbatim excerpts).
• 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). https://doi.org/10.3102/0013189X013006004. Opened 2026-10-26. File: bloom1984_2sigma.pdf (Avanti archive).
• Cepeda, N. J., Vul, E., Rohrer, D., Wixted, J. T., and Pashler, H. (2008). Spacing effects in learning: A temporal ridgeline of optimal retention. Psychological Science, 19(11), 1095-1102. https://escholarship.org/uc/item/0kp5q19x. Opened 2026-10-26. File: cepeda2008_ridgeline.pdf (in-press version).
• Dunlosky, J., Rawson, K. A., Marsh, E. J., Nathan, M. J., and Willingham, D. T. (2013). Improving students’ learning with effective learning techniques. Psychological Science in the Public Interest, 14(1), 4-58. https://doi.org/10.1177/1529100612453266. Opened 2026-10-26. File: dunlosky2013_pspi.pdf (Avanti archive).
• Kalyuga, S. (2007). Expertise reversal effect and its implications for learner-tailored instruction. Educational Psychology Review, 19, 509-539. https://doi.org/10.1007/s10648-007-9054-3. Opened 2026-10-26 (copy hosted on a University of Kentucky course page). File: kalyuga2007_reversal_review.pdf.
• Kraft, M. A. (2020). Interpreting effect sizes of education interventions. Educational Researcher (EdWorkingPaper 19-10 copy). https://doi.org/10.3102/0013189X20912798; copy from https://edworkingpapers.com/ai19-10. Opened 2026-10-26. File: kraft2020_effectsizes.pdf.
• Krishna, K., Akyol, P., and Ozer, E. (2025). Do time-constraints matter? How, why, and for whom? NBER Working Paper 33679. https://www.nber.org/papers/w33679. Opened 2026-10-26. File: akyol_krishna2025_timeconstraints.pdf.
• Muralidharan, K., Singh, A., and Ganimian, A. J. (2019). Disrupting education? Experimental evidence on technology-aided instruction in India. American Economic Review, 109(4), 1426-1460 (working-paper copy archived by Avanti). https://econweb.ucsd.edu/~kamurali/papers/Working%20Papers/Disrupting%20Education%20(Current%20WP).pdf. Opened 2026-10-26. File: muralidharan2019_mindspark.pdf (Avanti archive).
• National Testing Agency (2026). Information Bulletin, Joint Entrance Examination (Main) 2027. https://cdnbbsr.s3waas.gov.in/s3f8e59f4b2fe7c5705bf878bbd494ccdf/uploads/2026/10/202610241890224107.pdf. Opened 2026-10-26. File: nta_jeemain2027_bulletin.pdf.
• National Testing Agency (2026). JEE (Main) B.Tech question paper, 2 April 2026, Shift 1. https://cdnbbsr.s3waas.gov.in/s3f8e59f4b2fe7c5705bf878bbd494ccdf/uploads/2026/04/202604092096865379.pdf. Opened 2026-10-26. File: nta_jeemain2026_apr02_s1_paper.pdf.
• Nickow, A., Oreopoulos, P., and 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. https://www.nber.org/papers/w27476. Opened 2026-10-26. File: nickow2020_tutoring.pdf (Avanti archive).
• Pekkarinen, T. (2014). Gender differences in strategic behaviour under competitive pressure: Evidence on omission patterns in university entrance examinations. IZA Discussion Paper 8018. https://docs.iza.org/dp8018.pdf. Opened 2026-10-26. File: pekkarinen2015_omission.pdf (Avanti archive).
• Roediger, H. L., and Karpicke, J. D. (2006). Test-enhanced learning: Taking memory tests improves long-term retention. Psychological Science, 17(3), 249-255. https://doi.org/10.1111/j.1467-9280.2006.01693.x. Opened 2026-10-26. File: roediger2006_testing.pdf (Avanti archive).
• Rohrer, D., Dedrick, R. F., and Stershic, S. (2015). Interleaved practice improves mathematics learning. Journal of Educational Psychology (advance online 2014). https://doi.org/10.1037/edu0000001. ERIC ED557355, https://files.eric.ed.gov/fulltext/ED557355.pdf. Opened 2026-10-26. File: rohrer2015_interleaved.pdf.
• Rohrer, D., Dedrick, R. F., Hartwig, M. K., and Cheung, C.-N. (2020). A randomized controlled trial of interleaved mathematics practice. Journal of Educational Psychology, 112(1), 40-52. ERIC ED595322, https://files.eric.ed.gov/fulltext/ED595322.pdf. Opened 2026-10-26. File: rohrer2020_interleaved_rct.pdf.
• 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. https://doi.org/10.1037/a0037559. Opened 2026-10-26. File: rowland2014_meta.pdf (Avanti archive).
• Sadler, P. M., and Sonnert, G. (2016). Understanding misconceptions: Teaching and learning in middle school physical science. American Educator, Spring 2016. ERIC EJ1094278, https://files.eric.ed.gov/fulltext/EJ1094278.pdf. Opened 2026-10-26. File: sadler2016_misconceptions_ameducator.pdf.
• Schroeder, N. L., and Kucera, A. C. (2022). Refutation text facilitates learning: A meta-analysis of between-subjects experiments. Educational Psychology Review, 34(2), 957-987. https://doi.org/10.1007/s10648-021-09656-z (Europe PMC PMC8784251). Opened 2026-10-26. File: schroeder2022_refutation_meta.xml.
• Slavin, R. E. (1987). Mastery learning reconsidered. Report No. 7, Center for Research on Elementary and Middle Schools (also published in Review of Educational Research, 1987). ERIC ED294891, https://files.eric.ed.gov/fulltext/ED294891.pdf. Opened 2026-10-26. File: slavin1987_mastery_reconsidered.pdf.
• van der Haar, L., Gray-Lobe, G., Kremer, M., and de Laat, J. (2023). The long-term distributional impacts of a full-year interleaving math program in Nigeria. EdWorkingPaper 23-876. https://doi.org/10.26300/t8ht-2r44. Opened 2026-10-26. File: vanderhaar2023_nigeria_interleaving.pdf.
• van Gog, T., and Sweller, J. (2015). Not new, but nearly forgotten: The testing effect decreases or even disappears as the complexity of learning materials increases. Educational Psychology Review, 27, 247-264. https://repub.eur.nl/pub/92178/REPUB_92178-OA.pdf. Opened 2026-10-26. File: vangog2015_testing_complexity.pdf.
• von Hippel, P. T. (2024). Two-sigma tutoring: Separating science fiction from science fact. Education Next, 24(2). https://www.educationnext.org/two-sigma-tutoring-separating-science-fiction-from-science-fact/. Opened 2026-10-26. File: vonhippel2024_twosigma.html.
• Wisniewski, B., Zierer, K., and Hattie, J. (2020). The power of feedback revisited: A meta-analysis of educational feedback research. Frontiers in Psychology, 10, 3087. https://doi.org/10.3389/fpsyg.2019.03087. Opened 2026-10-26. File: wisniewski2020_feedback.pdf.

Leads, not opened

These were not obtained in full, so their findings are not used above.

• Sinha, T., and Kapur, M. (2021). When problem solving followed by instruction works: Evidence for productive failure. Review of Educational Research, 91(5), 761-798. https://doi.org/10.3102/00346543211019105. CC BY on SAGE, but the site served a bot-verification page to curl and the browser; the ETH repository copy (hdl 20.500.11850/490417) returned errors. Only the ERIC abstract was seen.
• Kulik, C.-L. C., Kulik, J. A., and Bangert-Drowns, R. L. (1990). Effectiveness of mastery learning programs: A meta-analysis. Review of Educational Research, 60(2), 265-299. https://doi.org/10.3102/00346543060002265. Closed access (Unpaywall, Semantic Scholar); not on ERIC in full text.
• Kulik, J. A., and Kulik, C.-L. C. (1988). Timing of feedback and verbal learning. Review of Educational Research, 58(1), 79-97. https://doi.org/10.2307/1170349. Closed; only the ERIC abstract was seen.
• Mullet, H. G., Butler, A. C., Verdin, B., von Borries, R., and Marsh, E. J. (2014). Delaying feedback promotes transfer of knowledge despite student preferences to receive feedback immediately. Journal of Applied Research in Memory and Cognition, 3(3), 222-229. https://doi.org/10.1016/j.jarmac.2014.05.001. Closed (Unpaywall).
• Butler, A. C., Karpicke, J. D., and Roediger, H. L. (2007). The effect of type and timing of feedback on learning from multiple-choice tests. Journal of Experimental Psychology: Applied, 13(4), 273-281. https://doi.org/10.1037/1076-898X.13.4.273. No open copy found.
• Muller, D. A., Bewes, J., Sharma, M. D., and Reimann, P. (2008). Saying the wrong thing: Improving learning with multimedia by including misconceptions. Journal of Computer Assisted Learning, 24(2), 144-155. https://doi.org/10.1111/j.1365-2729.2007.00248.x. Closed; Sydney repository copies of the thesis (hdl 2123/1693) and a 2007 proceedings paper (openjournals.library.sydney.edu.au/IISME/article/view/6345) blocked automated download. Only Crossref metadata and the proceedings abstract were seen.
• Crouch, C. H., and Mazur, E. (2001). Peer instruction: Ten years of experience and results. American Journal of Physics, 69(9), 970-977. https://doi.org/10.1119/1.1374249. Closed; the ComPADRE copy requires login.
• Smith, M. K., et al. (2009). Why peer discussion improves student performance on in-class concept questions. Science, 323, 122-124. https://doi.org/10.1126/science.1165919. Closed.
• Kalyuga, S., Ayres, P., Chandler, P., and Sweller, J. (2003). The expertise reversal effect. Educational Psychologist, 38(1), 23-31. https://doi.org/10.1207/S15326985EP3801_4. UNSW record is metadata only. The 2007 review above was used instead.
• Riener, G., and Wagner, V. (2017). Shying away from demanding tasks? Experimental evidence on gender differences in answering multiple-choice questions. Economics of Education Review, 59, 43-62. https://doi.org/10.1016/j.econedurev.2017.06.005. Closed; no working-paper copy found.
• Brunmair, M., and Richter, T. (2019). Similarity matters: A meta-analysis of interleaved learning and its moderators. Psychological Bulletin, 145(11), 1029-1052. https://doi.org/10.1037/bul0000209. Closed. (Cited second-hand by van der Haar et al. 2023 as finding 0.34 SD for mathematics; not verified here.)
• Yang, C., et al. (2021). Testing (quizzing) boosts classroom learning: A systematic and meta-analytic review. Psychological Bulletin, 147(4), 399-435. https://doi.org/10.1037/bul0000309. Closed.
• Murray, E., Horner, A. J., and Göbel, S. M. (2025). A meta-analytic review of the effectiveness of spacing and retrieval practice for mathematics learning. Educational Psychology Review, 37, 75. https://doi.org/10.1007/s10648-025-10035-1. Paywalled; only the abstract was seen. Worth obtaining: it is the most recent maths-specific synthesis.
• Danielson, R. W., et al. (2025). The effectiveness of refutation text in confronting scientific misconceptions: A meta-analysis. Educational Psychologist. https://doi.org/10.1080/00461520.2024.2365628. Closed.