How Studying Works | Latent Context Learning — How the Mind Learns Conflicting Rules Without Being Told the Context Changed

HSW-0221

A learner can know two rules and still fail because the harder job is deciding which rule is active now. The same word can mean something different in another sentence. The same symbol can carry a different role in another branch of mathematics. The same movement can need a different correction when the tool, surface or goal changes. Sometimes the context is announced. Often it is not.

That creates a study problem deeper than memorising more facts. If two situations share many of the same cues but demand different responses, the learner has to infer a hidden state: what kind of situation am I in? Only then can the right memory, rule or prediction be expressed.

The direct answer is this: people can sometimes learn conflicting associations without being explicitly told when the context has changed. Recent experimental work shows that learners can extract enough structure from recent experience to make context-dependent predictions even when the relevant context is latent. But this is not a licence to hide instructional structure from students. It is evidence that learning includes a contextual-inference problem, and that good studying should train learners not only to know a rule but to recognise the conditions under which that rule should be used.

The hidden question underneath many mistakes

Suppose a student has learned two procedures. Procedure A works in one family of problems. Procedure B works in another. In a worksheet where every A problem is grouped together and every B problem is grouped together, performance may look excellent. The page itself is supplying the context.

Now mix the problems and remove the headings. Nothing about the procedures has changed, yet a new task appears before either procedure can begin: identify the active problem class. A learner who cannot infer that context may execute a remembered procedure perfectly and still be wrong.

This is why the sentence “I know how to do it” is incomplete. A usable capability often has at least two layers:

Content or procedure: what to retrieve or execute.
Context discrimination: when that content or procedure applies.

The second layer becomes especially important when contexts overlap. If every context has a unique colour, heading or location, discrimination is easy. If the same cues appear in several contexts and only their recent pattern reveals which one is active, the learner has to infer more.

What a 2026 experiment actually tested

In an open-access 2026 iScience study, Fleming C. Peck, Hongjing Lu and Jesse Rissman examined whether people could learn conflicting visual associations when context changed over time. Their article, “Spontaneous emergence of context-dependent statistical learning in humans and neural networks”, used two human experiments with separate groups of 50 participants.

Participants watched a long stream of object images. Their overt task was simple: identify a small symbol embedded in each object. Unknown to them, the image sequence contained repeated pair relationships. The difficult part was that most of those relationships were context-dependent. An object that predicted one follower in Context A could predict a different follower in Context B.

In the first experiment, context switches were not explicitly signalled. In the second, a visual border signalled context. After learning, participants completed forced-choice tests in which recent sequence history established a context and they had to predict what should come next.

The important result was not perfect learning. It was that participants performed above chance on context-dependent associations, including direct-conflict trials where the alternative answer was the correct continuation in the other context. The authors also reported that the explicit border cue did not produce a clear context-dependent accuracy advantage over the unsignalled condition in this particular task. Most participants showed little explicit awareness of the temporal pair structure.

That result narrows the scientific claim. It shows that, under these laboratory conditions, recent sequence history contained enough information for people to acquire and express some conflicting context-dependent associations without an announced context label. It does not show that students learn school subjects best when teachers conceal the structure, nor that explicit explanation is unnecessary.

Context is often something the learner has to infer

A useful theoretical background comes from James Heald, Máté Lengyel and Daniel Wolpert’s review, “Contextual inference in learning and memory”. Their framework treats context not simply as a visible label but as a latent cause that may have to be inferred from feedback, sensory cues, time and recent experience. A related review, “The Computational and Neural Bases of Context-Dependent Learning”, develops the same problem across domains.

This distinction matters because two different events can look similar from the outside. Sometimes learning changes because the learner has built or updated a memory. Sometimes performance changes because a different existing memory is being expressed after the learner infers a different context. Those are not the same mechanism.

For studying, that means a sudden failure does not always imply that knowledge vanished. The learner may be selecting the wrong knowledge for the inferred situation.

A worked example: two algebraic rules that compete

Consider an illustrative learner who has practised expanding brackets and factorising expressions. In blocked practice, the instruction “Expand” appears above ten consecutive questions. Later, “Factorise” appears above ten more. The learner becomes fluent.

Now remove the command word and ask the learner to transform an expression into the most useful form for the next step. The learner stalls. This does not prove that either procedure was forgotten. The failure could be contextual: the learner has not learned which structural cues make expansion useful and which make factorisation useful.

A stronger practice sequence would therefore expose the learner to competing possibilities and require a classification before execution:

What is the current goal?
Which features of the expression matter?
Which transformation makes the next operation easier?
What alternative method is tempting here, and why is it less suitable?

The educational proposal here is an application, not a direct finding from the visual-sequence experiment. The experiment demonstrates latent context learning in a controlled task; the algebra example uses that mechanism as a disciplined design question.

Why conflicting rules are harder than two separate rules

If Cue X always predicts Y, accumulating experience is comparatively simple. If Cue X predicts Y in one context and Z in another, the cue by itself is insufficient. The system needs additional information that disambiguates the active relationship.

This produces three possible learning outcomes. First, one association can dominate and interfere with the other. Second, the learner can average or blur the associations and become unreliable in both contexts. Third, the learner can separate the contexts well enough to preserve both relationships and switch between them.

The third outcome is what flexible learning requires. It is also why contextual cues have to be chosen carefully. A cue that happens to correlate with the answer during practice may become a crutch rather than part of the real decision structure. If every geometry question of one type is printed on blue paper, blue may predict the method while the actual geometric relationship remains poorly learned.

The neural-network result is a model, not a brain scan

Peck and colleagues also trained recurrent neural networks on the same associative structure. Some network configurations developed distributed internal states that separated contexts and handled conflicting predictions better than others. The modelling was useful because it showed one computational way latent context representations can emerge from sequential experience.

That should not be converted into the claim that the human brain used the same hidden units, weight settings or algorithm. The networks are explanatory models. Similar behavioural patterns can make a model worth testing without proving biological identity.

This is an important evidence discipline for learners too. A model can clarify what must be computed without establishing exactly how a person computes it.

What this changes about study design

The simplest practical consequence is to add a context decision before the answer. Many study systems begin with “solve this.” A stronger system sometimes begins with “what kind of situation is this, and what evidence tells you?”

This can be trained in several ways.

1. Contrast near-neighbours

Place two cases beside each other that share surface features but require different treatment. Ask for the smallest feature that changes the correct response. The comparison teaches a boundary, not merely two answers.

2. Remove decorative cues gradually

Headings, chapter labels and teacher prompts can be useful during acquisition. They should not remain the only context signal if the learner will later face mixed or unlabeled questions. Fade them and see whether task-relevant structure takes over.

3. Ask for a rule plus its operating conditions

A flashcard that says “What is the rule?” tests content. A second card asking “When should this rule not be used?” tests the boundary. Conditional knowledge is part of usable knowledge.

4. Mix only after the components exist

Context discrimination cannot rescue a procedure that was never learned. Early blocked practice can reduce unnecessary load for a novice. Mixed practice becomes valuable when the next learning job is method selection or contextual discrimination rather than first exposure.

5. Test context after a delay

Immediate performance can be supported by the recent lesson, page layout or tutor’s wording. Return later with fresh examples, shuffled order and fewer labels. If the learner still selects the correct rule and can explain the cue, the contextual knowledge is more likely to be independently available.

A diagnostic for “I knew it but used the wrong thing”

When a learner applies a valid method in the wrong situation, do not immediately reteach the method. First ask where the selection failed.

Could the learner execute the correct method when explicitly named?
Could the learner distinguish this case from a close competitor?
Which cue did the learner use to classify the problem?
Was that cue causally relevant, or merely correlated with the answer during practice?
Would the learner make the same choice if formatting, order and wording changed?
Can the learner state both the trigger for the rule and a boundary case where it fails?

If execution is strong but classification is weak, the repair target is context discrimination. More repetitions of the already fluent procedure may add little.

Where this appears outside school

The same general problem appears in training and professional work. A technician may know several troubleshooting routines but must infer which system state is active from noisy signals. A clinician may know several diagnostic pathways but must determine which evidence is discriminating rather than merely common. A pilot may have procedures that depend on aircraft state, weather and phase of flight. A programmer may know several patterns but must identify which constraints make one architecture suitable.

These examples illustrate the structure of the problem. They are not evidence that the 2026 visual statistical-learning result directly transfers to each profession. The common question is whether multiple stored responses have to be selected under uncertain context.

The dangerous shortcut: teaching the label instead of the context

A learner can become excellent at responding to chapter titles. “This is the simultaneous-equations chapter, so I use simultaneous equations.” “This is the inference worksheet, so every question must require inference.” “We are revising acids, so every reaction must be interpreted through acids.”

That is not useless. Labels can organise early learning. But if the final environment is not labelled, the learner eventually has to recover the classification from the case itself.

A useful tutor therefore watches for prompt inheritance: the student succeeds because the lesson supplies the decision that the student will later need to make alone.

For parents and tutors: teach the fork in the road

When two methods are easily confused, the most valuable explanation may be the fork between them. Instead of another full demonstration, place two short cases together and ask: “What makes these different?” Then make the learner name the cue before solving.

Keep the language concrete. “Use this because the unknown appears in both equations” is better than “this is a Type 4 question.” “Use this tense because the later past event is being located relative to an earlier past event” is better than “this is the purple worksheet.” The objective is to move the context signal from the teaching environment into the structure of the task.

Then remove yourself. Give several fresh cases, including tempting near-neighbours, and require the student to classify them without hints. Ask for the reason for the classification, not just the final answer.

What the research does not justify

The 2026 experiments used young adults, novel visual objects and temporal pair statistics. The test followed the learning phase. They did not compare school curricula, long-term examination outcomes, children of different ages or explicit conceptual instruction. Above-chance context learning was not perfect context learning. The lack of a benefit from the simple border cue in that paradigm does not imply that meaningful instructional labels are useless.

The neural-network component should also remain in its proper category: computational modelling. It helps generate mechanistic hypotheses. It does not establish that human learners use an identical architecture or parameter regime.

These limits make the educational conclusion more precise, not less useful. Learners can be sensitive to hidden contextual structure, but teaching should still ask whether the available cues are the ones learners will need later.

An independent performance check

After teaching two or more competing rules, wait. Then build a short test with four properties:

examples are new rather than copied from instruction;
problem types are mixed;
chapter labels and method prompts are removed;
the learner must name the deciding cue before executing the answer.

Score selection and execution separately. A learner who chooses correctly but calculates badly needs a different repair from one who calculates beautifully with the wrong method. That distinction turns “careless mistake” into something diagnosable.

The larger lesson

Studying is not only the accumulation of answers. It is also the construction of a control system that decides which answer, rule, representation or action belongs to the present situation.

When contexts are obvious, that control problem disappears into the background. When contexts overlap, it becomes part of expertise. The learner must detect what changed, infer which situation is active, retrieve the corresponding knowledge and remain ready to switch when the evidence changes again.

That is why a learner can possess two correct rules and still need more learning. The missing knowledge may be the invisible sentence between them: use this one here, for this reason.