State of the art, per sub-question
Two records are kept apart: most adopted (what textbooks, sites and teachers use) and best evidence (what a study measured). They rarely have the same holder.
Records per sub-question
| Sub-question | Most adopted | Best evidence | Measured how |
|---|---|---|---|
| Q1 order of types in a set | blocked sets with a short mixed review: 9.7% of 13,505 textbook problems interleaved (Rohrer 2020b) | Rohrer 2020 (core): cluster trial, d = 0.83; Rohrer 2015: d = 0.79 at 30 days | unannounced delayed tests of mixed problems, classroom randomisation |
| Q1 order across sets | one lesson per set; Saxon's daily cumulative practice as the exception (Rohrer 2020b) | Murray 2025: spaced against massed in mathematics, g = 0.28, 0.24 inside courses (secondary) | meta-analysis of 27 studies |
| Q2 example to problem | textbook: worked examples in the lesson, then a block of problems | example-problem pairs over problems only (van Gog 2011, linked); fading over pairs (Renkl 2002; Atkinson 2003); adaptive fading over fixed (Salden 2009) | randomised experiments, near and far transfer |
| Q2 completion and Parsons | Parsons problems in e-books (Runestone, CSEd 2017d); fading examples into problems endorsed (Pashler 2007) | Paas 1992 for worked and completion problems; Ericson 2018 for Parsons: as effective as writing code, more efficient (abstract), with replication lacking (Du 2020) | transfer tests; time to solve |
| Q3 difficulty | starred or graded problems (MathEd 2014f); problems that open harder ones (Project Euler 2026, example only) | Jansen 2013b: higher preset success rate, more practice, larger gain (secondary) | adaptive practice, six weeks, 207 children |
| Q3 variability | NCETM variation sets (NCETM 2020) | Paas 1994 (core); Likourezos 2019: helps the experienced, hurts novices (secondary) | transfer tests |
| Q4 hints | on-demand hints in online homework (Roschelle 2016; ASSISTments 2026, example only) | Roll 2011 changed help use; Aleven 2016 no domain gain; Bastani 2025 hint-only tutor avoided harm (secondary) | classroom trials in tutors; field experiment, about 1,000 students |
| Q4 solutions | odd-numbered answers at the back (OpenStax 2026); solutions manuals; publish after submission (MathEd 2017a) | Rystrom 2005, two cohorts: lower scores with the manual (secondary) | course quizzes and exams |
| Q4 copying | sites selling solutions: Chegg 3.6 million subscribers at the end of 2024 (Chegg 2025) | Palazzo 2010: copying detected in logs went with a 2 SD decline in relative score on similar exam problems (secondary) | log analysis of an online tutor |
| Q5 volume | about 15 exercises per section, three sections a week, in one calculus class (MathEd 2014g, example only) | Rohrer 2006: 9 problems no better than 3 in one session (secondary) | delayed tests at one and four weeks |
| Q5 online homework | WeBWorK (since 1994; WeBWorK 2026), ASSISTments (owner's claim, example only) | Roschelle 2016: g = 0.18, 2,850 students, 43 schools | randomised field trial, end-of-year standardised test |
Adoption counts, descriptive, read 2026-10-29
| Count | Value | Caveat |
|---|---|---|
| interleaved problems in six representative US mathematics texts | 9.7% of 13,505 (Rohrer 2020b) | the authors' choice of texts |
| interleaved problems in the Saxon grade 7 text | 3,016 of 4,507 (Rohrer 2020b) | one text |
| mixed review in textbooks surveyed informally | under 15% of practice problems (Rohrer 2015) | informal survey |
| OpenStax Calculus volume 1, chapter 1 | answers printed for odd-numbered exercises 1 to 335 only (OpenStax 2026) | one chapter read |
| WeBWorK Open Problem Library | over 35,000 problems (WeBWorK 2026) | project's own page; no user count given |
| ASSISTments | over 1 million students since 2019; over 200,000 problems (ASSISTments 2026) | owner's claim, example only (7 of 20); its "73% more growth" cites no study |
| Project Euler | 151,219 members who solved at least 25 problems (Project Euler 2026) | undated count, example only (9 of 20) |
| Chegg subscription services | 3.6 million in the fourth quarter of 2024, down 21%; 6.6 million in 2024 (Chegg 2025) | not a count of homework copying |
| Classroom studies found by the EEF review | 12 on interleaving, 22 on worked examples (Perry 2021) | studies run in schools only |
Upper bounds (achieved)
- Interleaving whole homework sets raised delayed unannounced test scores in middle-school classrooms, d = 0.79 to 1.05 (Rohrer 2015; Rohrer 2014, secondary) and d = 0.83 in a cluster trial (Rohrer 2020, core).
- Online homework with hints and teacher reports raised an end-of-year standardised test (TerraNova), g = 0.18 (Roschelle 2016).
- Backward fading with principle prompts gave medium to large transfer without extra time (Atkinson 2003, secondary).
- A hint-only generated tutor kept practice gains without the later loss that open access caused (Bastani 2025, secondary).
Lower bounds (shown not to work, or to cost)
- Interleaving over a full year in 62 Nigerian classrooms: +0.29 SD on short-term tests, no average effect on the cumulative test (van der Haar 2023, secondary).
- Mixed against blocked in a university calculus course: no difference on the final (Gorgievski 2024, secondary).
- More problems of one type in a session: no gain at one or four weeks (Rohrer 2006, secondary); 50% more per set: no gain (Gorgievski 2024).
- Online against paper homework: no difference over 18 studies (Bishop 2024, secondary).
- Help-seeking training: better help use, no better learning (Aleven 2016).
- Open access to a model during practice: 17% lower exam grades afterwards (Bastani 2025, secondary).
- Order of examples and problems: no effect for inquiry skills (Ganaiem 2025); none at test with identical items (van Gog 2011b, secondary).
Holder of the record
- No study holds a record for a whole exercise set built from all the parts at once: opening pairs, fading, a short block, interleaved review, hints before solutions. Each part has its own holder.