Skip to content

Training and Measuring Cross-Domain Transfer: The Design of Abstractopia

A foundational methods paper.

Companion instrument to the Encyclopedia of Abstractions (Abstractopedia). This paper describes the design and the falsifiable claims of Abstractopia, a spaced-practice application for training and measuring the recognition of shared abstract structure across surface-dissimilar domains. It is a design-and-hypotheses contribution: the efficacy claims below are stated as predictions to be tested, not as results already in hand. Where we cite prior findings we have verified each reference against its primary source.

Abstract

The central skill of expertise, insight, and analogical reasoning is the ability to see the structure a situation shares with a distant one, independent of the surface features in which that structure is dressed. This ability is also the classic failure point of education: skills learned in one context notoriously fail to transfer to another. Abstractopia is an attempt to attack that failure directly. It treats cross-domain structural recognition as a trainable meta-skill and builds the training on three commitments: (1) teach structure, never surface; (2) measure transfer by requiring recognition of the same abstraction in an ever-farther domain, with a justification of why; and (3) teach the limits of each abstraction — where it breaks — as explicitly as the abstraction itself. The content is organized around a curated set of substrate-neutral "primes" drawn from the Encyclopedia of Abstractions. A late module reframes cognitive biases — themselves abstractions, some prime and some domain-specific — not as defects but as failure modes of good reasoning tools — a tool used on ground it does not fit — and trains the learner to read that fit, with a mastery gate that structurally refuses to certify a reflexive "bias-spotter." We describe the pedagogy, the assessment model, and the item-design discipline that gives the assessment its validity, situate the design against the transfer and learning-science literatures, state the specific predictions it makes, and note honestly the substantial risks — chief among them that far transfer is famously hard to produce and that a well-motivated design is not the same as a demonstrated result.

1. The problem worth attacking

Ask why an expert physicist solves a problem a novice cannot, and the interesting answer is not that the expert knows more formulas. When Chi, Feltovich, and Glaser asked experts and novices to sort physics problems, the novices sorted by surface — pulleys with pulleys, inclined planes with inclined planes — while the experts sorted by deep principle: "these are all conservation-of-energy problems," across wildly different apparatus [1]. The expert sees the structure through the surface. That capacity — to recognize that two things sharing no surface features share a form — is what Gentner's structure-mapping theory formalizes: an analogy is a mapping of a system of relations from one domain to another, carried by the relations rather than the objects, and constrained by the higher-order systematicity of those relations [2].

The trouble is that this capacity is exactly what education has the hardest time producing. Gick and Holyoak's radiation problem is the canonical demonstration. Given the problem cold, about 10% of people find the "converging forces" solution. Given a story about a general taking a fortress with converging troops — a perfect structural analog in a completely different surface — and then the radiation problem, the rate rises to only about 30%, because most people do not spontaneously notice that the story is relevant. Given two analogs first, so that a schema can be abstracted from them, and a hint to use it, the solution rate climbs to about 52% [3]. The lesson is twofold and it is the seed of this entire project: spontaneous transfer is rare because retrieval is surface-bound, and it improves when the learner induces the shared schema from multiple, surface-varied instances.

The pessimistic reading of the broader literature reinforces this. Barnett and Ceci, reviewing decades of transfer studies, argued that the field's confusion about whether "far transfer" even happens stems from a failure to say along which dimensions a transfer is far; their taxonomy lays out nine such dimensions (knowledge domain, physical context, temporal context, functional context, social context, modality, and so on) and shows that the farther the transfer along more of them, the rarer success becomes [4]. The honest baseline, then, is skeptical: teaching a general "thinking skill" and expecting it to travel is the modal failure of educational interventions, not their modal success.

Abstractopia takes that skepticism as its design brief rather than as a reason not to try. If transfer is rare because retrieval is surface-bound, then the training should relentlessly decouple retrieval from surface. If transfer improves when a schema is induced from multiple surface-varied instances, then the core loop should be exactly that induction, run to saturation and spaced over time. And if we are going to claim transfer, we should measure it the hard way — by demanding recognition in domains progressively farther from where the pattern was learned.

2. The substrate: substrate-neutral primes

The content rests on the Encyclopedia of Abstractions (Abstractopedia), a curated corpus of roughly 1,300 "primes" — recurring structural patterns each defined by a substrate-neutral signature (roles and the relations among them) that travels across domains — organized into a hierarchy of prerequisite and composition relations. Feedback, constraint, trade-off, incentive, signal-and-noise, path dependence, selection, regression to the mean: each is specified not by any home domain but by the relational skeleton that any instance must instantiate. Crucially, the corpus is not limited to such "positive" patterns: it also contains their failure-mode shadows — the characteristic way a pattern misfires when used off the ground it fits — and domain-specific abstractions whose vocabulary does not travel. This is the region where, as §3.4 develops, cognitive biases live: a bias is not a separate kind of thing from a prime, but an abstraction of a particular sort. The Encyclopedia is the theory; Abstractopia is one runtime for it — an instrument that teaches a curated subset of keystone primes and, in doing so, tests whether the theory's central promise (that these patterns are learnable as transferable structure) holds for human learners.

Two design consequences follow immediately. First, because a prime is defined by structure and not surface, every prime affords an unbounded supply of surface-distinct instances — which is exactly the raw material the Gick–Holyoak result says schema induction needs. Second, because the corpus is an external, human-curated answer key, "which prime is this an instance of?" has a ground truth independent of any individual item's author — a property that becomes important both for assessment validity here and, later, for using the same instrument to benchmark machines (§8).

3. The method

3.1 Teach the structure, never the surface

The first commitment turns Gentner's theory into pedagogy. A pattern is never introduced by definition first; it is introduced by comparison. The learner meets two vignettes that share a structure and share no surface — deliberately, no common domain, characters, or setting — and is asked what is true of both that mentions the specifics of neither. The zero-shared-surface constraint is load-bearing: comparison across similar surfaces teaches a domain; comparison across maximally dissimilar surfaces forces the relational skeleton into the open. This is schema induction by construction, and it mirrors the condition under which Gick and Holyoak saw transfer improve.

3.2 Measure transfer directly — the assessment is the contribution

Most "critical thinking" products cannot tell you whether they worked, because they never define the target behavior operationally. Abstractopia's assessment model is an attempt to make transfer measurable without cheapening it. Three pieces do the work.

Surface distance as the difficulty axis. Every practice item is tagged with its surface distance from the pattern's home domain — near (everyday, close), mid (moderately distant), far (a different far corner of the world each time: nature, history, cooking, physics, music, machines, the body). Mastery is defined not as "answered many items" but as "recognized the pattern far from home, and explained why." Transfer is thus not assumed; it is the very thing the top mastery state requires the learner to demonstrate.

The anti-flashcard invariant. No two items may share a surface. Reusing a surface would let the learner build a surface→answer association — the exact failure mode we are trying to train out. Enforcing surface uniqueness means the only stable thing across a pattern's items is its structure, so the only learnable regularity is the structure. This is the operational heart of "teach structure, not surface."

Justification as a gate. At far distance a correct label with a wrong why does not count. To reach the "Owned" state the learner must, after a built-in delay, repeatedly recognize the pattern in far domains and name the load-bearing relation that carries it, and must reject its tempting look-alikes. Requiring the justification turns the assessment from "can you pick the right word" into "can you articulate the structure," which is the construct we actually care about and the one least gameable by surface cues.

Together these make a concrete, falsifiable operationalization of transfer: a transfer profile — accuracy as a function of surface distance — and a transfer gap, the difference between near and far accuracy. A learner (or, later, a model) who has genuinely internalized the structure has a small gap; a surface pattern-matcher has a large one.

3.3 Teach the limits — where the tool breaks

The third commitment is the one most educational content avoids: teaching the edge of an abstraction. An abstraction is a projection — it keeps some relations and discards the rest — so two cases that share it are guaranteed to match on what it kept and to differ everywhere it was silent. Push an analogy past the shared projection into that discarded residue and it breaks, and the breaking point is locatable in advance: it is the edge of the shared structure. Abstractopia teaches this directly (a "failure-of-extension" drill: given a sound transfer, name the dimension along which it first breaks) because the twin of recognizing a pattern is knowing where the recognition stops paying. A learner who can spot patterns but cannot find their edge transfers confidently and is wrong in ways they cannot see — arguably a worse state than not transferring at all.

3.4 Cognitive biases as failure modes of good tools

A clarification a fresh reader should not have to infer: a cognitive bias is itself an abstraction — a point in the same structural space as everything else in this paper. Like any abstraction in the corpus, a bias is either a prime (substrate-neutral and cross-domain — Goodhart's law, regression to the mean[5], and selection bias are all biases that are genuine primes) or domain-specific (its vocabulary bound to one home domain, as with base-rate neglect). What makes a bias distinctive is not that it is a different kind of thing from the reasoning tools the app teaches, but where it sits relative to them: a bias is the failure-mode abstraction — the shadow — of a reasoning-tool abstraction, defined against the sound use of that same tool on ground it fits. (The sound-use case is thus not a separate abstraction but the very same tool on matching ground; the shadow is defined by contrast with it.) So the app does not change subject when it reaches biases; it points the same machinery at a particular region of abstraction-space.

The most distinctive module extends §3.3 to the learner's own reasoning. The conventional framing of cognitive biases[6] — the mind as a buggy machine, biases as the bugs — is both contestable and pedagogically corrosive: it manufactures the cynic who wields "that's just your bias" as a rhetorical weapon while remaining blind to their own. The ecological-rationality tradition offers a better frame: a heuristic is not good or bad in the abstract but is ecologically rational to the degree it is adapted to the structure of its environment [7]. Base-rate neglect is the shadow of trusting strong specific evidence — a normally excellent move that misfires precisely when the thing tested for is rare and the test is imperfect, which is exactly when natural-frequency representations restore correct reasoning [8]. "Confirmation bias" is better understood as a generally sensible positive-test strategy that misfires when the test is non-diagnostic [9]. So a bias is not a defect of the machine; it is a good tool used on ground it does not fit.

Abstractopia therefore teaches each bias as the failure mode of a reasoning tool the learner already owns, and unifies them under a single cross-cutting abstraction we call fit-relative competence: a tool's value is a relation to its environment, so the very structure that makes it a strength in a matching environment is what makes it a characteristic failure in a mismatched one. The trained skill is not "spot the bias" but "read the ground" — is this a situation where the tool is trustworthy, or one where it will mislead? Every misfire item is paired with sound-use twins in which the same heuristic is correctly applied, so that answering "misfire" reflexively is itself a scored error.

This design choice is also an ethical one made structural rather than exhortative. The mastery meter tracks success on misfire items and on sound-use twins separately, and certification requires competence on both. A learner who cries "bias" at every quick judgment accumulates no credit toward mastery — the schema literally cannot certify a reflexive bias-caller. The motivation for building the humility in as a gate rather than a slogan is the bias blind spot: people reliably see bias in others far more than in themselves, and merely being told about a bias does not dislodge the belief that one is personally exempt [10]. If the blind spot operates in the first person, the training must operate there too — which is why the app's productive surfaces ask the learner to catch misfires in their own cases, not only in tidy strangers' examples.

3.5 The engine: spaced, interleaved, effortful retrieval

Underneath the pedagogy is a scheduler built on the most robust findings in learning science. Practice is retrieval, not review, because repeated testing produces substantially greater long-term retention than repeated study — the testing effect [11]. Practice is distributed, not massed, because the spacing effect is one of the most replicated results in the field, with the optimal gap widening as the desired retention interval lengthens [12]. Patterns are interleaved rather than blocked, and the difficulty is deliberately desirable — the conditions that feel harder in the moment (spacing, retrieval, interleaving, varied surfaces) are the ones that build durable, transferable learning, even though they depress immediate performance and thus feel worse to the learner [13]. The surface-distance ladder is itself a desirable difficulty: it makes each encounter harder in exactly the way that, the theory predicts, builds transfer.

3.6 Item-design validity — the anti-gaming discipline

An assessment that claims to measure structural recognition is only valid if its items cannot be solved by surface shortcuts. In building the content we repeatedly discovered — and had to eliminate — ways an item could be answered without engaging the structure: the correct option being reliably the longest; a confidence word in the vignette predicting the answer; a detector's false-positive rate being stated only when the answer was "misfire"; the emotional tone or scenario type correlating with the verdict. Each of these is a construct-irrelevant cue that would let a learner (or a machine) score well while the assessment silently measured the wrong thing. We now enforce a battery of automated checks — option-length balance, phrasing balance, self-containment, absence of scenario/affect confounds, and numeric consistency — as a gate on every item. We flag this not as housekeeping but as a methodological point: for a transfer assessment, adversarial item validity is not optional; it is the property that makes the transfer-gap metric mean what it claims. The same discipline is what will let the instrument benchmark machines honestly (§8), where surface-cue exploitation is, if anything, the larger risk.

4. Why the mechanism is plausible

None of the above proves Abstractopia works. But two bodies of evidence make its mechanism more than a hope.

First, the app's core loop is a scaled, spaced version of the one manipulation that reliably produced transfer in the lab. Gick and Holyoak's jump from 30% to 52% came precisely from inducing a schema across multiple surface-varied analogs and then cueing its use [3]. Abstractopia does this not once but continuously, across an unbounded supply of surface-distinct instances, spaced and interleaved. If schema induction across varied surfaces is what moves transfer, the app is an attempt to run that intervention to saturation.

Second, and more pointedly for the biases module: the single debiasing approach with the strongest evidence is essentially the app's method. Morewedge and colleagues found that a single training intervention — a game or video that had people recognize bias-triggering situations across many varied scenarios with immediate feedback — produced medium-to-large, persistent reductions in several biases, outperforming the usual approach of explaining the bias [14]. Crucially, a follow-up showed those effects transfer to the field: trained graduate students were 29% less likely to choose an inferior, hypothesis-confirming solution on an unannounced business case modeled on the Challenger decision [15]. Situation-recognition trained by spaced, interleaved, feedback-rich, varied exposure is exactly what Abstractopia is — which means the app is not betting against the debiasing literature but building on its most encouraging result.

5. What we predict — the grand experiment

Because the whole point is to measure transfer, the design commits to specific, falsifiable predictions. A proper evaluation should be pre-registered; the core hypotheses are:

  1. Transfer gap narrows with training. Within a pattern, far-distance accuracy will rise relative to near-distance accuracy over the training arc — i.e., the transfer gap shrinks — more than a matched control that practices only near items. (A design that trained surface associations would show the opposite: near mastery with a persistent far deficit.)
  2. Untrained-domain transfer. On held-out far domains never seen in training, trained learners will recognize the pattern above a matched control — the direct test of transfer in Barnett–Ceci's strong sense.
  3. Justification predicts durability. Learners whose far-distance justifications name the load-bearing relation (versus a surface cue) will show better retention at a delay, linking the articulated schema to durable transfer.
  4. Bias training reduces misfires without inflating false alarms. After the biases module, learners will correctly flag more genuine misfires on novel situations and will not over-flag sound uses — the sound-use twins let us measure both, and the humility gate predicts the false-alarm rate should not climb (the failure mode of naive debiasing).
  5. Edge-finding is separable and trainable. Performance on failure-of-extension items will dissociate from recognition performance and improve with practice — evidence that "knowing where a pattern breaks" is a distinct, trainable competence, not a byproduct of recognition.

Any of these can fail. Prediction (2) is the one the transfer literature says is hardest, and a clean null there would itself be informative: it would locate precisely where a well-motivated, well-instrumented attempt at trainable transfer runs into the same wall that has stopped others. The instrument is designed so that failure is legible rather than ambiguous.

6. Limitations and honest risks

We hold the design to the same standard it teaches — name where it breaks.

  • Far transfer is hard, and good design is not evidence. The base rate of success for "learn to think" interventions is low [4]. Alignment with the mechanisms that have worked (§4) raises the odds; it does not settle them. Until the experiments in §5 are run, the efficacy claims are hypotheses.
  • It is demanding. The desirable difficulties that build transfer feel worse in the moment [13]; a learner's intuitive sense of progress will lag their actual learning, which is a real engagement and retention risk at scale.
  • Quality is gated on curation. The assessment's validity depends on the corpus's quality and on the anti-gaming discipline (§3.6). A weak prime or a leaky item degrades the measure; this is why the item-validity gate is treated as load-bearing rather than cosmetic.
  • Selection effects in any early study. The learners most drawn to this already think this way; demonstrating transfer to that population is not the same as demonstrating it creates the capacity in a general one. Sampling has to be designed against this survivorship trap — itself, fittingly, one of the primes the app teaches.
  • Bias content wears empirical caveats. The biases literature has a replication problem; the module includes only biases whose structure survives even if particular lab effect sizes shrink (mathematical facts, sampling facts, or field-replicated effects), and it teaches the honest edges (e.g., that base-rate misfires largely vanish under natural-frequency formats [8]).

7. A note on machines

The same instrument that measures structural transfer in people could, in principle, probe it in language models. Present a vignette in one domain, ask which prime it instantiates, then test whether the model recognizes the same prime in a domain sharing zero surface — the transfer gap, computed for a model. Because the correct label is fixed by the external corpus rather than by the item's author, and because the anti-gaming discipline strips the surface cues a model would otherwise exploit, such a measure would resist the usual failure of "reasoning" benchmarks, where models score well by latching onto artifacts. That humans and models could take the same instrument, and that their transfer profiles could be compared, is a research affordance worth noting — though building and validating it is work for another day.

8. Conclusion

Abstractopia is a wager that the most valuable and least-taught cognitive skill — seeing shared structure across the surface, and knowing where that structure stops — can be trained with the best-established machinery of learning science, if the training is disciplined about attacking surface at every turn and honest about measuring the result the hard way. The wager may not pay off; far transfer has humbled better-funded attempts than this one. But the design is coherent end to end: a corpus of substrate-neutral structure, a pedagogy that induces that structure from surface-varied instances, an assessment that defines mastery as transfer and refuses to let surface stand in for it, a module that turns the learner's own reasoning failures into the same lesson, and a set of falsifiable predictions that make success and failure equally legible. If it works, it is a rare thing: a trainer for the skill that lets everything else transfer. If it fails, it will fail informatively — and that, too, is worth building.

References

[1] Chi, Michelene T. H., Paul J. Feltovich, and Robert Glaser. "Categorization and Representation of Physics Problems by Experts and Novices." Cognitive Science 5, no. 2 (1981): 121–152. Experts sort problems by deep principle where novices sort by surface features — the empirical seed of the structure-over-surface thesis. ↩

[2] Gentner, Dedre. "Structure-Mapping: A Theoretical Framework for Analogy." Cognitive Science 7, no. 2 (1983): 155–170. The structure-mapping account: a sound analogy maps a system of relations, not surface attributes — the theoretical basis for "load-bearing vs. surface." ↩

[3] Gick, Mary L., and Keith J. Holyoak. "Schema Induction and Analogical Transfer." Cognitive Psychology 15, no. 1 (1983): 1–38. Comparing two surface-different analogs induces a schema and roughly doubles spontaneous transfer — the single best-supported result the app's core loop scales. ↩

[4] Barnett, Susan M., and Stephen J. Ceci. "When and Where Do We Apply What We Learn? A Taxonomy for Far Transfer." Psychological Bulletin 128, no. 4 (2002): 612–637. The canonical taxonomy of transfer distance — and the sober base rate for "learn to think" programs — that this design holds itself to. ↩

[5] Galton, Francis. "Regression towards Mediocrity in Hereditary Stature." Journal of the Anthropological Institute of Great Britain and Ireland 15 (1886): 246–263. The original account of regression to the mean — an extreme measurement tends to be followed by a less extreme one, by the statistics alone. ↩

[6] Tversky, Amos, and Daniel Kahneman. "Judgment under Uncertainty: Heuristics and Biases." Science 185, no. 4157 (1974): 1124–1131. Founds the heuristics-and-biases program — judgment leans on a few efficient heuristics that produce systematic errors — the tradition the biases module reframes as tool-fit. ↩

[7] Gigerenzer, Gerd, Peter M. Todd, and the ABC Research Group. Simple Heuristics That Make Us Smart. New York: Oxford University Press, 1999. Argues heuristics are ecologically rational — sound or unsound relative to the environment they meet — the "fit to environment" frame the biases module adopts. (No stable DOI.) ↩

[8] Gigerenzer, Gerd, and Ulrich Hoffrage. "How to Improve Bayesian Reasoning without Instruction: Frequency Formats." Psychological Review 102, no. 4 (1995): 684–704. Base-rate "neglect" largely dissolves when problems are posed in natural frequencies — the honest edge taught alongside the base-rate bias. ↩

[9] Klayman, Joshua, and Young-Won Ha. "Confirmation, Disconfirmation, and Information in Hypothesis Testing." Psychological Review 94, no. 2 (1987): 211–228. Reframes "confirmation bias" as a generally adaptive positive-test strategy that misfires only in certain structures — the nuance the app teaches instead of a blanket rule. ↩

[10] Pronin, Emily, Daniel Y. Lin, and Lee Ross. "The Bias Blind Spot: Perceptions of Bias in Self versus Others." Personality and Social Psychology Bulletin 28, no. 3 (2002): 369–381. People see bias readily in others but not themselves — why explaining a bias rarely dislodges the belief that one is personally exempt. ↩

[11] Roediger, Henry L., and Jeffrey D. Karpicke. "Test-Enhanced Learning: Taking Memory Tests Improves Long-Term Retention." Psychological Science 17, no. 3 (2006): 249–255. The testing effect — retrieval practice beats restudy for durable retention — the basis for recognition-first, effortful items. ↩

[12] Cepeda, Nicholas J., Harold Pashler, Edward Vul, John T. Wixted, and Doug Rohrer. "Distributed Practice in Verbal Recall Tasks: A Review and Quantitative Synthesis." Psychological Bulletin 132, no. 3 (2006): 354–380. Meta-analytic evidence for the spacing effect — distributed practice reliably beats massed practice for retention — the basis for the spaced spiral. ↩

[13] Bjork, Elizabeth L., and Robert A. Bjork. "Making Things Hard on Yourself, but in a Good Way: Creating Desirable Difficulties to Enhance Learning." In Psychology and the Real World, edited by Morton A. Gernsbacher et al., 56–64. New York: Worth Publishers, 2011. Introduces "desirable difficulties" — conditions that feel harder and slow apparent progress yet build durable, transferable learning. (Book chapter; no stable DOI.) ↩

[14] Morewedge, Carey K., Haewon Yoon, Irene Scopelliti, Carl W. Symborski, James H. Korris, and Karim S. Kassam. "Debiasing Decisions: Improved Decision Making with a Single Training Intervention." Policy Insights from the Behavioral and Brain Sciences 2, no. 1 (2015): 129–140. A single situation-recognition training — spotting bias-triggering situations across varied scenarios with feedback — produced large, persistent debiasing, outperforming explanation. ↩

[15] Sellier, Anne-Laure, Irene Scopelliti, and Carey K. Morewedge. "Debiasing Training Transfers to Improve Decision Making in the Field." Psychological Science 30, no. 9 (2019): 1371–1379. The situation-recognition debiasing effect transfers out of the lab — trained students made better decisions on an unannounced, unrelated real case. ↩