Interleaving Table

Interleaving is practising the same material from more than one direction, a way to learn where it begins and ends. A cube makes this visible. Each face is a different pattern, the knowledge is the cube in the hand, and one line of questioning never shows the whole thing.

The Interleaving Table is a menu of ways to test from memory, meant to be raided rather than read. Two or three methods, picked by whether the knowledge is explainable or executable, and a short list of things that feel like practice and are not. Spaced Interleaved Retrieval is the parent loop and owns the spacing schedule these angles get spread across: when a session happens, how the delays lengthen, and what the session hands back.

Core takeaways

• Interleaving is practising the same material from more than one direction. A subject switch inside one sitting is one form of it; shifting the angle inside one subject needs no switch at all.
• The benefit is largest where items are similar but clearly distinct. On isolated word lists blocked study wins, and expository text shows no reliable effect.
• Volume, order, and type are settings on a session, each with its own test: how much retrieval before a hole appears, whether the demand relates two or more concepts, and whether an error is one of approach or of pure execution.
• Open wide and high-order about two days after learning, log the detail gaps, and narrow over weeks. A session that finds many gaps is the session working.
• Methods are picked by whether the knowledge is explainable or executable and by the height of the demand. Two or three run properly beat eight sampled.
• Passive rereading, repetition past mastery, practice questions answered only in the head, and undirected group discussion are the things to stop.

The narrow version, and where the mix gets dropped

The version most rooms already own is a subject switch inside one sitting: biology then maths then biology again. That still counts. It is one face of the cube, not the cube. Angle-shifting needs no subject switch. The same subject holds: move between its concepts, compare them, return, challenge each from a new direction. Mixing problem types inside one subject, not mixing subjects, is the form that has been measured in real classes: about six in ten items correct against under four in ten after blocked practice, at one month.

A first pass that maps relations is already mixed, because comparison is the work of encoding. The system’s position is that a first pass done right is interleaved, full stop. Retrieval is where the mix gets dropped. The same one or two testing methods get used every time, and the drop is easy to miss. That is why a spaced retrieval system has to enforce the mix explicitly, and why this page lives under retrieval rather than encoding. The retrieval load is heaviest exactly when encoding is weakest.

Similar, but clearly two things

Similar but clearly distinct is the window.

Probing two confusable concepts from several angles shows exactly where the boundary runs, and memory of each deepens. That is the mechanism the comparison research actually supports. Push the pair too close and the effect often dies: if a listener would treat the two names as one thing, varied testing has nothing left to separate. Push them too far apart and the effect dies for the opposite reason. The difference is already obvious, so contrasting them defines no edge.

Interleaving is not a free upgrade on every material. If the items do not need telling apart, mix less. Across a large set of studies the benefit sits around a moderate size, largest where categories are similar but distinct, paintings more than mathematics, and it reverses on isolated word lists, where blocked study wins. Expository text shows no reliable effect. A vocabulary or word-list session that mixes by default is using the tool on the material that does not want it.

Varying the topic while holding the angle fixed is not interleaving at all. The window is a property of the angle, not of the timetable.

Three dials on a session

Volume, order, and type are settings on a session, not three new techniques. Each is defined by its test.

Volume is how much ground the session covers, measured by the number needed to find a gap. A full brain dump is high-volume. A targeted session probes two or three concepts. High volume pays when gaps are everywhere; targeted pays once they are rare. Volume is a dial on a method, not a different method: teaching two or three related concepts is the low-volume setting of the same technique whose high-volume setting is teaching the whole topic.

This system's own picture, not a finding: a ten-hour dump across
thirty pages of writing can surface five gaps. The number needed
to find a gap is the volume reading.

Order is how integrated the demand is. High-order tasks make concepts interact; low-order tasks retrieve isolated facts. The test: does answering require relating two or more concepts, or one in isolation? The full form adds a qualifier: the influence those concepts have on each other is part of what must be worked out. Inside the high-order band there is a spectrum. Ranking three concepts by which matters most is high order. Holding eight and their interrelations is much further up. A reader who treats order as a switch runs one masterclass and stops.

Type is the mix of explainable knowledge and automatic execution. Explainable knowledge is facts, concepts, logic. Executable knowledge is skills and processes, sequences and decisions in context. Most real work needs both, in different mixes. Match methods to whether the subject is mostly concepts that can be said, or mostly execution that has become automatic. Type is a distribution to match, not a side to pick: a heavily conceptual subject runs mostly declarative with a little procedural; mathematics carries a large conceptual share against a real procedural one; the retrieval mix should mirror the subject’s mix. Using facts to work up a case and produce a plan is still declarative. Procedural means execution that has become automatic. Declarative, Procedural, and Conditional Knowledge owns those definitions, including why applying knowledge is not procedural, and including the conditional kind the executable half is building.

One hard challenge classifies the error. A miss in approach, strategy, or connecting logic is declarative. A miss in pure execution is procedural: the right equation with wrong arithmetic; sound architecture with buggy syntax; a strong argument in clumsy prose. The dial exists because the session is there to find gaps, and a session testing the wrong kind of knowledge will not find them. A faultless declarative pass on an execution-bound subject reports no problem.

Even in execution-heavy subjects the usual bottleneck is declarative: the approach, the strategy, the choice of move. Those decisions carry the consequences, which is why they resist being handed off. Interleaving for Complex Problem Solving extends the same multi-angle move into knowledge work and messy problems.

Dial	The test
Volume	How much retrieval before a hole appears
Order	One concept in isolation, or two or more and their influence
Type	A fault of approach, or a fault of pure execution

The angles themselves

Each angle rebuilds the same knowledge by a different route. Say it, use it, sketch it, or throw it at a problem whose shape was never practised: each route forces a different rebuild. The differences between those rebuilds are the schema, and that is what a curveball actually tests.

A working menu, not a curriculum:

• Teach an imaginary student. A full out-loud lesson; every clumsy explanation is a located gap. A real friend fills the gaps.
• Modify a hard question. Change its variables or add a step, so the answer can no longer be pattern-matched.
• Frame it in the real world. Write a test question that drops the material into a context where it would actually be used.
• Draw instead of write. Finding a visual form is itself a fresh angle.
• Quiz and debate. In a study group, quiz each other; for suitable subjects, stage a debate where one side argues the other has missed a point.
• Build or apply it. Take the knowledge into something made, or into a problem that arrives with its own context. This is the menu item that tests execution rather than explanation.

No combination is universally best. Subject and context decide, and that takes experimentation. Two or three good methods that follow the trend are enough; the practice becomes competent quickly once it is actually run.

What to stop

These are not the default. For the same time there is always a better alternative.

Passive rereading, recopying, and relistening do not activate the pathways that make learning stick. They are slow and demotivating. Rereading is the measured case; recopying and relistening sit in the same family.

Repetition past mastery is the next exclusion: REBIM, repeating a skill already done to a high level, at the same difficulty. Same scripts, same-difficulty maths problems, the same conversation when already fluent. Six extra same-difficulty problems after the first success produced no gain at one week and none at four, while the same practice split across two sessions a week apart produced no gain at one week and a large gain at four. Same effort, different placement, opposite result. It feels productive because smooth performance is indistinguishable from progress: learners rated blocked practice as more effective right after a test they had just done better on under interleaving.

Practice questions used passively, answering only in the head or checking before a full attempt, fail on both halves. Ability is overestimated and challenge is underestimated, and practice questions are a finite resource, so a wasted one is gone. Answers get produced in full, written or typed, never only mentally.

Undirected group discussion underperforms a protocol. Structured small-group work can raise achievement; most group study spends more time than the same result costs elsewhere. The protocol lives on Group Study.

Testing explainable knowledge

Science, law, theory-heavy finance, vocabulary and grammar sit here. Isolated fact, a relationship between two or three ideas, then a judgment about which relationship matters more: that is the ladder. A handful of methods work at more than one height.

Height	What it asks	Methods that sit here
Lower	Direct recall, isolated definitions	Cover-copy-check; simple cards; linear dump; isolated questions; isolated teaching
Mid	Apply it, obvious relationships	Relational cards; map dump; relational questions; teaching; direct practice questions; the exchange protocol
Higher	Unfamiliar contexts where several things interact	Closed-book map rebuild; evaluative cards and questions; map dump; whole-part-whole; directed discussion; extended questions; the exchange protocol

Cover-copy-check: cover part of the notes, rebuild it from memory, check. It is evidenced for short-term retention of basic information, mostly with school-age children. Image-occlusion cards are the practical variation. Best on labelled diagrams and processes.

Simple cards test isolated facts. When the encoding underneath is weak they multiply: a climbing card count is a signal about the encoding, not a reason to study harder. Steep forgetting and wasted repetition follow. Flashcards owns the restriction these rows assume: what one card may hold, and what a climbing count says.

Write evaluative questions in one session, answer them in another. Both halves are retrieval events, so splitting them buys two, and the writing session sets up the answering one. Writing those questions is itself a retrieval event; handing the job to a model is a time trade that buys an outside check. Both are real; neither is the default.

Brain dumps split by form, and the order is the opposite of instinct. The map form goes early: it finds the big gaps fast, better at higher order, worse at fine detail. The linear form goes late, just before an assessment, because it finds gaps only by dumping until something stumbles.

Isolated teaching is slow for what it returns, which is why cards usually win at that layer. Micro-retrieval is the useful form: teach the piece just processed, inside the same session. Whole-part-whole reteaching is the hardest method here. The benefit tracks skill at the technique, so it returns little until the encoding under it is strong. If the reteach feels easy, it is not happening. WPW owns that reteach and the diagnostic.

Rebuilding an old topic as a closed-book map is a re-encoding move, so it belongs first in a cycle with time left after it. Struggling to build one is the diagnostic that the material was never processed that deeply, not a sign of being bad at maps.

Explaining in plain terms as if to someone with no background returns a moderate teaching benefit across a large set of studies. That measurement is prompted self-explanation inside instruction, so the method is micro-retrieval, not a system.

Extended practice questions, four steps:
1. Answer from memory.
2. Mark every place that felt unsure.
3. Build a perfect answer sheet from the material.
4. Check it against the official answers.
Uncertainty counts as a gap even when the answer was right,
because one right answer does not survive a variation.
A large block, or split across sessions.

Testing with feedback substantially outperforms testing without it; elaborated retrieval is one of the factors that decide whether retrieval transfers at all.

The exchange protocol, for anyone who can arrange one person at the same level or above: each writes an exam of mid and higher-order questions and builds a perfect answer sheet for it; swap exams and answer from memory; mark uncertainty; build a sheet for the exam received; swap sheets; compare and argue the differences. The generative work is in the writing, answering, sheet-building and comparing, six of the eleven steps, not in the discussion. Split across days.

Mnemonics, drawn scenes and story links, stay for ordered lists and discrete volumes of isolated information. Learned early they become a substitute for encoding. Six weeks of loci training roughly doubled word-list recall, a real effect confined to exactly that narrow demand. Method of Loci owns the drawn-scene variant. Rote Learning and Memorisation is the parent of those rows and of the bound on when memorising is the right move.

Testing executable skill

Retrieved execution, simple: practise the skill from memory with no cue, one isolated application, no variation. It is the right first move. It is also the exact thing that becomes repetition past mastery a week later, once the skill is comfortable.

Integrative: chain components into something larger. Basic functions chained into a complex one; simple phrases chained into a dialogue. The purpose is conditional knowledge, knowing when and how to use what is known. Move here as soon as basic competence exists, because it generates more, loads more, and interleaves more.

Applied: start from a target output and work backwards. Integrative builds up from components; applied works back from a product. An application someone actually needs, built end to end. A real scene in the language being learned, its plausible variations practised, then the scene entered for real. Realistic transfer does not follow automatically from realistic practice.

Challenges attach a prompt or problem stem to that execution: simple, integrative, or edge-case. Edge cases are unfamiliar by design and need strong explainable knowledge just to choose an approach.

Creating an edge case is far cheaper than solving one. Creating needs only the knowledge that two concepts could interact; solving needs the approach and the execution. Two consequences: edge cases can be built for material not yet mastered, and they can stand in for evaluative questions inside the exchange protocol. That is the join between the two halves of the table.

Variable modification is the same problem with new figures and the same approach; that is the mixed-problem-type practice measured in real classes. Variable addition is different: the learning is in choosing the variable, whether it is coherent, what value is reasonable, whether a solution still exists, how it interacts with what is already there. Air resistance and a collision changing a trajectory. A cylinder’s volume maximised under a surface-area constraint with radius to height fixed at 1:2. A timeout on a function. A conversation line grown into a conditional with a fallback.

Every method in the menu is the same act at a chosen height: produce it from memory, at an order that is currently uncomfortable, and read what failed. The order dial is the same dial on both sides of the split; the type only changes what “produce it” means. Strong explainable understanding supports executable skill and the reverse; most real tasks want both.

A menu invites collecting methods. Two or three run properly beat eight sampled, and time spent choosing is time not spent retrieving. If a method has become comfortable, it has stopped testing anything: raise the order or change the format. That is the same failure as repetition past mastery, arriving quietly.

Wide first, then narrow

Open high-order and high-volume, soon after learning. Around two days after first contact, this system’s own default rather than a measured optimum, teach the entire topic as a masterclass or rebuild it in a full structured dump. WPW is that reteach. Gaps are plentiful then, so an expensive wide pass finds many per hour. A narrow probe run first spends its time on ground that is already solid. The wide pass turns up a gap about every paragraph, in three recognisable forms: I do not remember that as well as I thought · I need the notes for the details · I cannot see how those two connect.

High-order gaps stay filled. A repaired relational gap is integrated knowledge, sticky, unlikely to reopen, which is why it earns the early, expensive treatment. Low-order detail behaves the opposite way. It is forgettable by nature and keeps needing return, which is why cards run in parallel rather than waiting. During the wide pass, log detail gaps without chasing them. The list becomes the agenda for a cheap targeted session later. A second route to a targeted session: open the later one with a test rather than working from the start of the topic until something breaks.

A second wide, high-order pass roughly a week later is still the early stage. The funnel narrows over weeks, not between sessions. After that, a session blends one harder challenge with the leftover facts that were logged. With the high-order scaffold in place, the tedious details finally have somewhere to attach. Learned first, they sit as disconnected fragments: nothing feels relevant, and there is no structure to file anything against, so the sense of what matters never engages.

For procedurally dominant material a teaching pass is the wrong opener however high-order it is. The early wide pass takes a procedurally dominant form instead.

Cards tick along beside the funnel from the early-to-mid stage onward, isolated-detail cards as a companion rather than a substitute.

Skip what can be looked up. If a detail never needs to come from memory in real use, drop it from retrieval and spend the time at mid-to-high orders.

Blocked practice feels better while interleaving works better. Messy is the expected feel, not the quit signal.

A session that finds few gaps while mastery is genuinely not high is the ineffective one. Many gaps is the session working. The honest way to have fewer early gaps is better encoding, not a gentler session. The fix for a pass that returns nothing is the level, the spacing, and the encoding, in that order.

The wide pass is expensive by design. It is the wrong move on material whose items do not need telling apart, on a subject whose bottleneck is execution when the angles are all declarative, and before encoding has landed. Revision is the lightweight on-ramp for a reader not yet running retrieval sessions, or when the full funnel is more than the material deserves. Nobody has to wait for the table to start.

The back-and-forth between ideas, and between angles on the same idea, is the shared engine behind high-order learning. Interleaving is the deliberate practice of it.

Working with a model

A language model raises the efficiency of these methods when used well and damages them when leaned on. That is a caution, not a finding.

Use it for the tedious low-effort work, keyword lists and a big-picture summary, for pulling many sources into one place, for generating questions, and for asking a topic open. Consolidating sources reduces a real cost, split attention; a model is not required to do it. One notes page does the same job. The questions clause is the same resolution as above: writing is a retrieval event, a model is a time trade with an outside check.

Avoid it for judging whether a nuanced claim is true, since these systems have no inherent grasp of truth however fluent the answer, and for the mental-model, importance, and relationship work. That work is what builds the network.

Beginners do not outsource chunking and organising until the difference between passive and active work can be felt. If a model is used anyway, challenge what it gives and look for the alternative it did not offer.

Adaptations, as tool classes: three different cards on one concept; questions at several heights with the answers checked; a brain dump handed over for gap-finding.

The menu in use

On explainable-heavy material the two or three methods look like this over a couple of months, offered as a guide to adapt rather than a schedule to obey. Day one, while studying: teach each piece immediately after processing it. Days two to five: a map brain dump, and cards made from the gaps it exposes. Those cards tick along in pockets of time. Days seven to fourteen: relational and evaluative questions, plus their answer sheets. Around a month: collect the questions and run the exchange protocol. Around two months: directed discussion, whole-part-whole if it can be run, or the protocol again.

Executable-heavy material gets the short version: start low, keep escalating. Skills build on themselves so repetition is automatic; competence tracks practice, feedback and difficulty; and higher-order methods test the lower orders directly, which is not true on the explainable side.

Mixed material runs the explainable cadence until the explainable half is competent, then switches to executable methods and climbs. The more executable the topic, the shorter the explainable stretch: sometimes the first encoding plus one session.

Session quality beats timing precision. A schedule off by days or weeks still works if the sessions are good. A first map brain dump that exposes no gaps means the test is sitting below the level already reached; move up a height rather than repeating the dump.

How to practice this

1. About two days after first contact, teach the whole topic out loud or rebuild it as a full structured dump. Expect a gap about every paragraph. Few gaps while mastery is low means the test sits too low; move up a height.
2. During that wide pass, write down each detail gap without chasing it. The list is the agenda for a cheap targeted session later.
3. Write evaluative questions in one session and answer them in another. Each half is a retrieval event, so the split buys two.
4. Mix problem types inside one subject instead of running a block of one type. It will feel worse than blocked practice; that feel is expected, not the quit signal.
5. Produce every practice answer in full, written or typed, never only in the head. Mark every place that felt unsure, even where the answer was right; each mark is a gap.
6. When a method has become comfortable, raise the order or change the format. A comfortable method has stopped testing anything.

Open questions

• Whether the funnel actually beats mixing from the start, or only feels more efficient: nobody has compared them.
• Whether the too-similar edge of the window is real: the too-different edge has been measured, the too-similar one has not.

Sources

• Dunlosky, J., et al. (2013). Improving students’ learning with effective learning techniques. Psychological Science in the Public Interest. Rereading: low utility.
• Rohrer, D., & Taylor, K. (2006). The effects of overlearning and distributed practise on the retention of mathematics knowledge. Extra same-difficulty problems past first success: no gain at 1 or 4 weeks.
• Kornell, N., & Bjork, R. A. (2008). Learning concepts and categories: Is spacing the “enemy of induction”? Fluency illusion: blocked practice rated as more effective immediately after a test already done better under interleaving.
• Rowland, C. A. (2014). The effect of testing versus restudy on retention. Psychological Bulletin. Recall outperforms recognition; testing with feedback d = 0.73 vs without d = 0.39.
• Springer, L., Stanne, M. E., & Donovan, S. S. (1999). Effects of small-group learning on undergraduates in STEM. Review of Educational Research. d ≈ 0.51. Structured cooperative work can raise achievement; undirected discussion underperforms a protocol.
• Skinner, C. H., McLaughlin, T. F., & Logan, P. (1997). Cover, copy, and compare. Short-term retention of basic information, mostly with school-age children.
• Bisra, K., et al. (2018). Inducing self-explanation. Prompted self-explanation g = 0.55 across 64 reports.
• Rohrer, D., Dedrick, R. F., Hartwig, M. K., & Cheung, O. S. (2020). A randomized controlled trial of interleaved mathematics practice. Journal of Educational Psychology. 787 students, 54 classes; interleaved problem types 61% vs blocked 38%, d = 0.83 at one month.
• Pan, S. C., & Rickard, T. C. (2018). Transfer of test-enhanced learning. Elaborated retrieval as a transfer condition; realistic practice does not automatically transfer.
• Dresler, M., et al. (2017). Mnemonic training reshapes brain networks. Six weeks of loci training roughly doubled word-list recall.
• Brunmair, M., & Richter, T. (2019). Similarity matters: A meta-analysis of interleaved learning and its moderators. Psychological Bulletin. 59 studies, 238 effect sizes, overall g = 0.42. Paintings ≈ 0.67, mathematics ≈ 0.34, expository text non-significant, words g = −0.39 (blocking wins).
• Carvalho, P. F., & Goldstone, R. L. The discrimination account of interleaving: the effect tracks similarity that needs discriminating.
• Fiorella, L., & Mayer, R. E. (2013, 2014). Learning by teaching. Teaching and explaining as a retrieval angle.