Exercise sets: how a set of exercises is ordered and supported

Question: how is a set of exercises ordered and supported, so that working through it builds a method the learner can apply to a new problem? And what does the evidence say about sequence, fading, hints and solutions?

It is a multivocal literature review (Garousi 2019) by one reviewer, searched on 2026-10-29.
It holds 78 new formal sources and links 13 more already reviewed in the core repository.
It scores 39 grey sources out of 20 (10 from the web and the What Works Clearinghouse, 29 Q/A threads); 12 fall below the threshold of 10 and are used as examples only.
It is not a standard, a new result, or a claim to the state of the art. A number read only in an abstract is marked "(secondary)".

Files

File	Role
protocol	how the pool was built: questions, strings with ERIC and Semantic Scholar counts, stopping rule, scoring
sources	the flow of the pool, then one row per source with its quality
frame	step 1: the problem in one sentence, the terms, what research and practice call it
review	step 2: existing reviews, seminal work, one strand per sub-question, practice, Q/A sites
state of the art	step 3: most adopted against best evidence, adoption counts, upper and lower bounds
position	step 4: known, not known, where practice and evidence part
against	null classroom results, contested orders, the Saxon case, the limits
baseline	step 5: the blocked sheet the shapes must beat, and how to measure it
shapes and checks	proposals for the blocks of a set and the checks a program could run
bibliography	every cited work: key, citation, link
handoff	works met here that belong to the sibling reviews
`flow.json`, `openalex.tsv`	the flow as data; the formal strings in OpenAlex syntax, for counting by script

Claims

Claim	Grade	Where
Interleaving the problem types of homework raised delayed classroom test scores: d = 0.79 at 30 days, 1.05 at two weeks	measured, classroom trials in mathematics (one read in full, one secondary)	review.md 2
A full year of interleaved sets raised short-term scores (0.29 SD) and not a cumulative test	measured, one cluster trial, working paper (secondary)	against.md
Spacing a type across sets beats massing it in mathematics, less inside courses (g = 0.28; 0.24)	measured, meta-analysis (secondary)	review.md 0
9.7% of 13,505 problems in six mathematics textbooks are interleaved	measured, descriptive	review.md 7
Most students judge the least spaced, least interleaved schedules most effective	measured, descriptive (secondary)	review.md 2
Nine problems of one type in one session did no better than three at one and four weeks	measured, one study (secondary)	review.md 6
Fading worked steps beats example-problem pairs on near transfer; adaptive fading beats fixed fading	measured (secondary)	review.md 3
Example-problem order beats problem-example in some studies and not in others	measured, contested	review.md 3; against.md
Parsons problems are as effective as writing the equivalent code, and faster	measured, one research group (abstracts)	review.md 3
Variability helps experienced learners and hurts novices	measured, one paper (secondary)	review.md 4
Training help seeking changed help use, not learning	measured (secondary)	review.md 5
Open model access during practice lowered later exam grades by 17%; a hint-only tutor avoided most of the loss	measured, one field experiment (secondary)	review.md 5
Homework copying rises toward deadlines and goes with lower exam scores	measured, descriptive (secondary)	review.md 5
Online homework with hints and teacher reports raised an end-of-year standardised test, g = 0.18	measured, one randomised field trial (read in full)	review.md 6
Online and paper homework do not differ	measured, meta-analysis of 18 studies (secondary)	review.md 0
Odd-numbered answers only, solutions released after submission, variation sets, starred problems	convention, untested	review.md 7, 8
A set built from all the measured parts beats the blocked textbook set	convention, untested (a hypothesis to test)	position.md; baseline.md

Named baseline

To beat: B1, the blocked sheet, one run of same-type problems per lesson with answers to half the items at the end (Rohrer 2020b; OpenStax 2026).
To match on the opening of a type: B2, the example-problem pair (van Gog 2011, linked; Pashler 2007).
Comparison: classes randomised, same problems, unannounced mixed tests at one and 30 days and at the end of the year; strategy choice, transfer, time and help use logged; d or g per outcome (baseline).

Linked, not reviewed again

Principles of the standard an exercise set serves: listed in the evidence README.
The effects themselves (interleaving, spacing, worked examples, fading): the core's learning science.