Practice has to teach me when

A whole page of correct answers can look like progress. Put “Independent events” above every problem, though, and the page is doing part of the thinking for me. I know which tool I'm expected to use. I want to know whether I can choose it when the heading is gone.

In the previous article, I described a teaching conversation that uses my reasoning to decide what to explain next. Practice should extend that conversation: bring an idea to mind, work out whether it fits, and use it to make a decision or investigate something. I also want tasks where deciding what needs solving is part of the work.

Bring the idea back before looking

Retrieval means bringing previously learned information to mind. For me, that can mean reconstructing an explanation, drawing a relationship, or trying a problem before opening the worked solution.

In a prose-learning experiment, repeated study produced better recall than repeated retrieval after five minutes. A week later, repeated retrieval performed better. Repeated study also increased confidence. These were short texts and recall tests, but the contrast makes a useful point: immediate fluency and later availability can tell different stories.

Retrieval can also be studied with questions that require reasoning. In experiments using science texts, retrieval combined with restudy outperformed concept mapping with the text visible on a one-week assessment, including inference questions. That comparison supports including reconstruction in conceptual learning; it leaves plenty of room for reading, examples, and other ways of building understanding.

For the counter example from the teaching article, I could put the solution away and explain why removing a red counter changes the next draw. I want to recover the relationship behind the calculation, then check it. If I cannot reconstruct it, that gives us a specific place to resume.

Make choosing part of the task

A retrieval cue is information that helps bring knowledge to mind. A chapter heading, a familiar diagram, or the resemblance to the previous exercise can all supply one. Those supports can help me learn a method. I also want tasks that ask me to decide which method applies.

Interleaving mixes problem types rather than grouping all problems of one type together. In a geometric-volume experiment, mixed practice produced lower practice accuracy but better test performance a week later than blocked practice. Only 18 students completed the experiment, and tutorial placement differed between conditions. The result is useful evidence for that arrangement, without isolating method selection as its sole cause.

In Make It Stick, Peter C. Brown, Henry L. Roediger III, and Mark A. McDaniel connect varied and interleaved practice with learning to distinguish kinds of problems and select an appropriate response. That is what I want to practise here: noticing which relationship matters before carrying out the calculation. (2014, Chapter 1)

My application would be a set of probability questions with different decisions to make: a single draw, two draws with replacement, and two draws without replacement. Before calculating, I would identify what can change between events and explain why the chosen calculation fits.

Some focused practice may still be useful while a method is unfamiliar. Once I can carry it out, I want opportunities to distinguish it from alternatives. Mixing cases should serve that decision, rather than become a rule to shuffle everything immediately.

Change the demand deliberately

I want to adjust practice according to the difficulty my attempt reveals. Familiarity, the number of ideas involved, and the help available give me different things to vary.

To change familiarity, keep the underlying relationship and change how the situation is described. Suppose a tray holds three faulty components and two sound ones. A quality check selects two at random, setting the first aside before selecting the second. What is the chance that both are faulty? The structure matches the counter problem, but the familiar bag and colour cues are gone.

To combine ideas, I would first make sure I have learned them separately. After learning conditional probability and complements, I could ask for the chance of at least one red when drawing twice without replacement from a bag of three red and two blue counters. A complement gives a route through the opposite event: subtract the chance of two blues from one. Here, that is 1 − (2/5 × 1/4) = 9/10.

To change support, I could keep a problem similar but leave out a prompt naming the method. If I get stuck, a question about what changes after the first selection may help me restart. I would record that help and later try a fresh case without it.

These are proposed practice choices. Changing one demand at a time would make it easier for me to inspect what caused difficulty. A problem that combines several unfamiliar ideas and removes every hint at once gives me much less specific information.

I also want practice through actual work. If I'm learning to evaluate a search system, I could put together test queries, define what a useful result would look like, and compare a proposed change with the current system. I'd inspect the failures as well as the score. If the change helps some queries and hurts others, I need to investigate why and decide what to try next. That gives the concepts a job to do, with an outcome I can examine.

Use the correction

An attempt gives us work to inspect. Did I choose the wrong relationship, forget a condition, or make an arithmetic error? Each suggests a different next step.

I want feedback to identify the relevant difference and let me revise the attempt. If I used 3/5 for the chance of another red after removing a red counter, we can examine the remaining counters. If I chose the correct conditional probability but multiplied incorrectly, we can work on that calculation. Another full explanation of the topic may miss the actual gap.

In actual work, feedback might be an unexpected output, a failed test, or someone finding my explanation unclear. I want to compare that with what I expected and work out what to investigate. A result going my way doesn't, by itself, show that my reasoning was right.

The information supplied during practice matters too. In a study of retrieval and transfer, adding focused key statements to rereading narrowed its disadvantage against retrieval with feedback, although retrieval still performed better. Time and spacing were not fully equated. The study helps qualify the comparison: access to the information needed for a later task was part of what differed.

My practical response is to check the reasoning against a dependable explanation and try again. If an AI-generated question is ambiguous or its correction looks wrong, we need to check that first. A flawed exercise gives us poor grounds for deciding what I understand.

Return after a gap

Spacing distributes learning across time. It gives me a way to check an idea after the immediate context of the lesson has faded.

In an experiment on learning short facts, the best gap before review increased as the final test moved further away. Increasing the gap helped up to a point and then hurt performance. That result supplies no universal timetable for every subject.

I would choose a later revisit with the goal in mind, then adjust according to what I can recover. Before displaying the old solution, I want another attempt. If I need a hint, that is useful information to keep. If the reasoning is missing altogether, we may need to return to teaching.

Spacing and interleaving ask different questions. I can revisit the same kind of problem after a delay, or mix several kinds within one session. A mixed set may also space repetitions of each kind, so the effects can overlap. I want to be clear about what I changed.

Be precise about what a new problem shows

Transfer means using prior learning in a changed task or situation. Renaming counters as components is a modest change of context. It gives me another opportunity to use the relationship; it cannot establish that I will spot it wherever it matters.

In Butler's cross-domain transfer experiment, retrieval with feedback outperformed repeated study on new inference questions a week later. Participants were explicitly told to use information from the earlier session, and each question mentioned the relevant concept. The result supports recall and application with those cues, while leaving spontaneous recognition a separate question.

That distinction should shape my progress notes. Solving after a hint, solving when the method is named, and recognizing and solving independently are different achievements. Returning successfully after a delay adds evidence about what remained available. I want the record to preserve those conditions.

AI can help propose the next task and respond to my reasoning. I want its choices to remain open to correction, and the problem and solution to remain checkable. Research informs the ingredients here; this combination is an approach to evaluate through actual work.

The next task should follow from what the last attempt revealed. Sometimes that means a less familiar application or less support. Sometimes it means revisiting a prerequisite with an explanation. I want practice to help me frame a problem, choose an approach, and check what happens when I use it. Books, documentation, and AI can be available; I still want to understand the decisions I make with them and revise those decisions when the evidence changes.