Spacing and interleaving

The same hours, arranged differently

Take six hours of revision for one topic. Spend them in one Sunday block and you will remember substantially less in three weeks than if you spend them as six one-hour sessions across three weeks. Nothing else changes. This is the spacing effect, and it is one of the oldest results in psychology, going back to Ebbinghaus in the 1880s and confirmed continuously since.

The mechanism is worth knowing because it makes the advice stick. Each time you retrieve something after partly forgetting it, the retrieval is effortful, and effortful retrieval strengthens the memory more. Study it again immediately and there is nothing to retrieve (it is still sitting in mind) so the repetition does almost no work. Forgetting a little is the ingredient.

That is why massed study feels so much better. Everything is fresh, nothing is hard, and the smoothness reads as competence. It is the fluency trap in a different costume.

Intervals that are good enough

People get lost trying to optimise the schedule. The gains from perfect intervals over roughly sensible ones are small. Roughly sensible:

• Same day, a few hours later.
• Two days later.
• A week later.
• Two to three weeks later.
• A month before the exam, and then once in the final week.

A rough guide that works: if the exam is n weeks away, the gaps between reviews can be around n divided by five. Nothing turns on the exact number. What turns on it is that a gap exists at all.

Anki implements this automatically and adjusts per card by how you rate the difficulty. If you do not want a system, a paper calendar with topic names on dates is entirely sufficient and takes ten minutes to build.

Interleaving: the harder one

Blocked practice is twenty problems of type A, then twenty of type B. Interleaved practice mixes them: A, C, B, A, C, B.

Interleaving produces worse performance during practice and better performance on a later test. In the classic demonstrations with mathematics problems, students practising in blocks outperform the interleaved group during the session by a wide margin and are then beaten by an equally wide margin on the test a week later.

The reason is specific and useful. In a block, you already know which method to use: the heading told you. On an exam paper, nobody tells you. Choosing the method is a separate skill from executing it, and blocked practice never exercises it. Interleaving is how you learn to recognise which kind of problem you are looking at, which is precisely what students who knew the material and still did badly were missing.

Getting a model to interleave

Models default to blocking, because a set of similar questions is the natural output of a prompt about one topic. You have to ask:

Here are the six topics from this module. Give me one question at a time, drawn at random from any of them, without telling me which topic it is from. After my answer, tell me the topic and whether I was right.

Without telling me which topic is the whole technique. That clause converts an exercise in execution into an exercise in recognition.

A harder variant for exam preparation:

Mix in occasional questions that look like topic A but are actually solved with topic B’s method.

Those are the ones that separate grades, and they are almost impossible to construct for yourself, because you know the answer.

What is not worth your time

Learning styles. The idea that you are a visual or auditory learner and should be taught accordingly has been tested repeatedly and does not hold up; matching instruction to a claimed style does not improve outcomes. Believe your preferences about what is pleasant, not about what works.

Highlighting. Consistently rated among the least effective techniques when studied. It feels like engagement and produces recognition, not recall. If you highlight, treat it as a way of marking what to later retrieve, never as the study itself.

Re-copying notes. Same problem. Copying is not encoding. Rewriting notes from memory is a different activity entirely and is excellent.

The scheduling conversation worth having

One good use of an assistant is planning:

I have five topics and three weeks. I can do 45 minutes a day, five days a week. Build a schedule that spaces each topic at increasing intervals and interleaves them within each session. Give it to me as a plain list by date.

The model is good at this because it is a constraint-satisfaction problem with a known answer shape. Check the arithmetic (it is the kind of thing it gets slightly wrong) and then follow it, which is the hard part and is not the model’s job.