How Known Concepts Become Layers of Structure
Principles and relationships can help people reason beyond familiar examples. Research on expertise, analogy, replay, and inference reveals both the promise and the limits of this layered account.
Part 3 of 4 in the series “The Background Work of Innovation.”
In the first two articles in this series, I examined two parts of a hypothesis about innovation: breaks and sleep can sometimes improve later problem solving, and existing knowledge shapes new ideas and learning. Neither result, by itself, explains how the mind represents a new problem.
This article examines the part of the hypothesis I find most interesting: known concepts expand into layers of structure, and those layers are what allow us to reason about something new.
The research does not identify a single mechanism that builds these layers. But several lines of evidence point in the same direction. Experts organize knowledge around principles rather than surface features.[1] Comparing examples can produce an abstract schema that transfers to new problems.[2] During rest, the brain can reorganize new experience according to structure it learned earlier.[3] And the ability to infer relationships that were never directly taught can emerge after an offline delay.[4]
Here, “layers” means relationships between particular cases, recurring patterns, and more general principles. It is an explanatory model, not a claim about literal anatomical layers or one required architecture of thought. A schema, an ordered list, a relational map, and a perceptual chunk are different kinds of structure. The question is when such representations help people reason beyond the examples they learned.
Experts organize knowledge by principle #
In a classic 1981 study in Cognitive Science, Michelene Chi, Paul Feltovich, and Robert Glaser asked eight advanced physics PhD students and eight undergraduates who had just completed a semester of mechanics to sort 24 physics problems by similarity of solution.[1]
The novices grouped problems by what they looked like, such as problems involving inclined planes or springs. The experts grouped problems by the principle needed to solve them, such as conservation of energy or Newton's second law.[1] The authors proposed that expert knowledge is organized around problem schemata, each containing the information needed to solve a category of problems.
This was a small comparison of existing expertise, not an experiment that randomly taught people one knowledge organization or another. It shows a difference in how these groups represented problems. It does not establish that organizing knowledge into layers caused expertise, or that every expert uses a single hierarchy.
It is also the clearest picture I know of what I mean by layers. An inclined plane is a known concept. Conservation of energy sits above it, connecting inclined planes, springs, pendulums, and cases the student has never seen. Experts do not just know more. Their knowledge is organized so that principles sit above cases.
Comparison builds a layer above the examples #
How does a layer like that get built? One answer comes from research on analogy.
In a 1983 study in Cognitive Psychology, Mary Gick and Keith Holyoak gave participants stories that illustrated a problem and its solution, then asked them to solve a different but analogous problem.[2] The target was a classic medical puzzle: how to destroy a tumor with radiation without destroying the healthy tissue around it.
When participants read only one story, attempts to help them extract its general principle, through summaries, a verbal statement of the principle, or a diagram, did not achieve notable success. When participants read two analogous stories and described how they were similar, they often derived a general problem schema as a byproduct. The quality of that schema strongly predicted whether they transferred the solution. The same verbal statements and diagrams that had failed with one story proved highly beneficial when paired with two.[2]
The comparison gave participants a way to notice a relationship shared across stories. It does not follow that one example cannot support transfer, or that any two examples will create an abstraction. Recognizing that a learned solution is relevant to a new problem remains part of the challenge.
This was laboratory research using stories and puzzles. But it offers a specific mechanism for my hypothesis: a layer of structure can form when known cases are compared and what they share becomes explicit.
The resting brain reorganizes experience #
Comparison can be deliberate. My hypothesis also proposes useful work outside active problem solving. A 2019 study in Cell by Yunzhe Liu, Raymond Dolan, Zeb Kurth-Nelson, and Timothy Behrens provides evidence about how prior structural knowledge organizes replay during rest.[3]
Participants learned a rule for recovering the correct order of objects from scrambled presentations. They later applied that rule to new objects. During a subsequent rest period, magnetoencephalography detected rapid sequences of activity representing the objects, a phenomenon known as replay.[3]
The replayed sequences did not simply repeat what participants had seen. They followed the order implied by the rule learned earlier. The replay also carried an abstract code for each object's position and sequence, which appeared about 50 milliseconds before the code for the object itself. The authors argue that this kind of representation helps a previously learned structure generalize to new objects.[3]
The distinction is crucial: replay expressed an organization based on a rule participants had already learned. The study did not show that rest invented the rule, built a new conceptual layer, or caused better reasoning. Replay strength also did not correlate significantly with behavioral performance in either experiment. Resting activity can reflect a learned structure without establishing when that structure was constructed.[3]
Structure makes new inferences possible #
Can people use learned relationships to infer a relationship they were never directly taught? And does a period away from learning improve that ability?
A 2007 study in PNAS by Jeffrey Ellenbogen and colleagues tested these questions. Fifty-six participants learned five pairwise choices that formed an unstated ordering, A over B through E over F. They were then tested after 20 minutes, 12 hours awake, 12 hours including sleep, or 24 hours.[4]
Recall of trained pairs was high in all groups. Accuracy on untrained pairings was about 52% after 20 minutes and above 75% after the longer delays. Across all inference pairs, however, the 12-hour sleep and wake groups did not differ significantly. The reported sleep advantage was selective: on B over E, accuracy was 93% after sleep and 69% after wakefulness.[4]
Confidence did not increase with the sleep-related advantage. The authors interpreted the result as stronger integration of the hierarchy, describing it as:
“a benefit that may operate below the level of conscious awareness.” [4]
Confidence and awareness are different measures. An unchanged confidence rating cannot establish that participants were unaware of the relevant relationships. Nor does performance after a delay reveal precisely when the representation formed. It might have developed during learning, between sessions, or while retrieving information at test.
Later results are mixed. Berens and Bird's 2022 study found no overnight benefit for inference accuracy in its task, although response times improved slightly. It also found neural signatures consistent with structural representations in some participants who inferred poorly. Structure alone was not sufficient for successful performance.[5]
Foldes and colleagues reported sleep-related inference benefits in 2023, with outcomes depending on learning strength and the positions of items in the hierarchy.[6] Taken together, these studies support a conditional relationship between learning, consolidation, and inference. They do not make an offline delay a universal prerequisite.
One more distinction matters for this series. B and E are distant positions within one ordered set. That is not the same as transferring an idea to a distant domain. Inferring an untrained relationship and inventing a solution to a new real-world problem are different achievements.
What a layer looks like in practice #
Return to the hypothetical operations engineer from the previous article. She has handled dozens of incidents involving retries, queues, and timeouts.
At the first level, she knows the incidents: what failed, when, and what fixed it. At the next level, she recognizes a pattern:
“Retries against an overloaded service make the overload worse.”
Above that sits a principle:
“Any mechanism that responds to stress by adding work can turn a small failure into a large one.”
When she designs a scheduler she has never built before, the principle is what she reasons with. She may conclude that retries need limits, that load should be shed before it spreads, and that a burst of failures should slow the system down rather than speed it up. None of those conclusions came from a single incident.
She may also struggle to explain how she knows. As I discussed in an earlier article, skilled performance and a complete verbal explanation are different things. One possibility is that she describes individual cases while relying on relationships among them. That is an interpretation of this hypothetical example, not something these experiments established about engineers.
The real test is a problem the layers were not built for #
Transfer would be evidence that a representation is useful beyond its training cases. I should not define a “layer” as whatever transfers successfully, because then the claim that layers improve transfer would be circular.
A fair test would define the proposed abstraction in advance and assess it independently of success on the final problem. It would randomly vary comparison practice and the interval before testing, while checking memory for the original cases. This would help separate the effects of deliberate comparison, elapsed time, and rest.
The test should include new examples with the same underlying relation, unfamiliar contexts where that relation still applies, and similar-looking examples where it does not. The last group matters: a useful abstraction needs boundaries as well as reach.
If my hypothesis is right, independently measured structural knowledge should predict appropriate transfer beyond what memory for cases predicts. That does not entail that the benefit must grow with every increase in domain distance. Failures on unfamiliar contexts or misleading lookalikes would help locate the limits of the representation.
Layers make knowledge usable #
The research supports several components of the hypothesis. Experts can organize problems around principles. Comparing examples can help people extract shared relationships. Resting replay can follow a learned structure. People can also infer untrained relationships. The stronger claim that offline processing builds conceptual layers which then cause new reasoning remains a hypothesis linking these findings.
In the final article in this series, I examine the complication: the same structures that make reasoning possible can also block the next idea.
The most important question is not “How much does someone know?” It is “What structure relates what they know, and what can they infer from it?”
Previous in this series: Part 2: Why New Ideas Are Built From Known Concepts
Next in this series: Part 4: Why Innovation Depends on Building Structure and Taking It Apart
References #
-
Chi, M. T. H., Feltovich, P. J., & Glaser, R. (1981). Categorization and representation of physics problems by experts and novices. Cognitive Science, 5(2), 121–152.
-
Gick, M. L., & Holyoak, K. J. (1983). Schema induction and analogical transfer. Cognitive Psychology, 15(1), 1–38.
-
Liu, Y., Dolan, R. J., Kurth-Nelson, Z., & Behrens, T. E. J. (2019). Human replay spontaneously reorganizes experience. Cell, 178(3), 640–652.e14.
-
Ellenbogen, J. M., Hu, P. T., Payne, J. D., Titone, D., & Walker, M. P. (2007). Human relational memory requires time and sleep. Proceedings of the National Academy of Sciences, 104(18), 7723–7728.
-
Berens, S. C., & Bird, C. M. (2022). Hippocampal and medial prefrontal cortices encode structural task representations following progressive and interleaved training schedules. PLOS Computational Biology, 18(10), e1010566.
-
Foldes, T., Santamaria, L., & Lewis, P. (2023). Sleep-related benefits to transitive inference are modulated by encoding strength and joint rank. Learning & Memory, 30(9), 201–211.
Cite this article #
When quoting or referencing this article, credit Paul Greenberg and RapidKT and link to the article. Include the section title when directing readers to a particular passage.
Suggested citation
Greenberg, P. (2026, October 1). How Known Concepts Become Layers of Structure. RapidKT. https://rapidkt.com/pages/blog/how_known_concepts_become_layers_of_structure
In-text reference: (Greenberg, 2026).