Skip to content

Analogical transfer and structural alignment

Part of Inverse Innovation with the Encyclopedia of Abstractions · Analogical transfer and structural alignment · Last revised August 2026

Review date: 2026-08-03
Scope: cognitive science and computational analogy, including structure-mapping, analogical access and encoding, far transfer, relational similarity, negative transfer, schema induction, case-based reasoning (CBR), derivational analogy, SME/MAC-FAC/LISA and successors, neural relational reasoning, and recent language-model benchmarks.
Claim discipline: this is a scoped evidence synthesis, not a historical-priority review. “Not located” means not found under this protocol and search scope; it does not establish nonexistence.

Executive conclusion

The literature strongly establishes the components from which an inverse-innovation system could be built, but it does not collapse them into one capability.

  1. Established: analogical mapping is not the same problem as analogical retrieval. When two situations are placed side by side, people and programs can prefer deep, systematic relational matches; when a reasoner must retrieve one situation from memory, surface and semantic overlap exert much stronger control. This dissociation is supported behaviorally by Gentner, Rattermann, and Forbus (1993) and computationally by MAC/FAC.
  2. Established: comparison, relational labeling, multiple varied examples, and contrasts with near misses can make relational structure more explicit and portable. The behavioral anchor is Gick and Holyoak (1983); computational demonstrations include SEQL, DORA, and ALIGN.
  3. Established: implemented systems demonstrate many individual operations: structural correspondence and candidate inference (SME), case-library retrieval (MAC/FAC), integrated access/mapping/inference and schema induction (LISA), relation learning (DORA), incremental relational generalization (SEQL/SAGE/ALIGN), case adaptation and revision (CBR), replay of problem-solving rationales and failures (PRODIGY), and procedure transfer across intermediate states (Proteus). These systems normally receive a target problem, a source–target pair, or an example stream. They do not thereby demonstrate open-ended discovery of target problems from a solution archetype.
  4. Established with important boundary conditions: neural and neuro-symbolic systems can learn or extract relational representations and generalize beyond familiar content, especially when training data, hard negatives, bottlenecks, or explicit graph structure force attention to relations. Yet performance is uneven under held-out relations, novel symbol systems, long contexts, and large-memory retrieval. Strong fixed-benchmark performance is not evidence of open-corpus analogical innovation.
  5. Nearest recent precedent: Shen, Druckmann, and Zou (2026), an unreviewed preprint with code and data, gives an LLM a biomedical research problem, extracts object mappings and shared relations, generates distant problem analogies, transfers solution methods back, conducts a literature-based novelty check, and implements one proposal for each of four biomedical problems. This is unusually close to an operational analogical-innovation pipeline. Its direction is still given target problem → distant source/problem → candidate solution, not given solution structure → search unrelated domains for target problems; the paper also identifies feasibility evaluation and end-to-end execution grounding as remaining bottlenecks.
  6. Provisional gap: in the peer-reviewed and artifact-bearing literature inspected here, I did not locate a reproducible system that begins with an explicit domain-general solution structure, systematically crosses it with multiple unrelated domains to formulate target problems, operationalizes complete solution proposals, conducts prior-art and usefulness/adopter scrutiny, retains negative cases and repair trajectories, and repeats the process as a matrix. This is a scoped gap statement, not a novelty or priority claim.
  7. Implication for EoA: the defensible contribution boundary is not “uses analogy,” “finds distant analogs,” “represents relations,” or “adapts cases.” Those are established. The testable contribution is the direction and integration of the workflow: explicit solution archetype → generated target problem → structural map → adapted proposal → falsifiable test → prior-art/usefulness critique → retained positive and negative trajectory, repeated across a declared domain matrix.

Terminology and task-direction distinctions

Term Operational meaning in this review Frequent source of confusion
Base/source Better-known situation from which structure or a procedure is projected. A source can be a solved case, a domain, a story, an image, or a generated exemplar; these are not equivalent evidence conditions.
Target Situation to be interpreted or solved. In most experiments and systems the target is supplied. “Target answer generation” is not “target problem generation.”
Surface similarity Overlap in objects, attributes, words, setting, or other readily cued content. Surface information is not always noise; it can constrain a useful mapping. The problem is reliance on it when it conflicts with causal/relational structure.
Relational or structural similarity Common relations and higher-order organization among roles, potentially despite different entities. Topic or embedding proximity is not by itself evidence of structural similarity.
Structural alignment / mapping Constructing consistent correspondences between roles in source and target, often preferring one-to-one, parallel, and systematic relational matches. Mapping presupposes that usable representations and candidate situations already exist.
Candidate inference / projection Transferring an unmapped source proposition or procedure to the target as a conjecture. A projected inference is a hypothesis, not a validated solution.
Access / retrieval / reminding Bringing a potentially useful source from long-term memory or a case collection into consideration. Good mapping once cued does not imply reliable retrieval from a large corpus.
Analogical encoding Comparing examples so their common relations become salient and encoded. This is a learning intervention, not merely presenting an analogy.
Schema induction Forming a more abstract relational description across cases. A schema can be descriptive; it is not automatically an actionable solution archetype.
Far transfer Transfer across substantial distance in domain, context, modality, or surface form. “Far” is multidimensional. A domain label alone does not measure structural or semantic distance.
Adaptation Modifying a retrieved solution, plan, or inference so it fits target constraints. Correct retrieval and mapping can still fail at adaptation.
Negative transfer Prior knowledge or a retrieved case worsens performance, commonly because surface overlap licenses a structurally inappropriate procedure. A rejected analogy is not necessarily negative transfer; the analogy must misdirect processing or outcome.
Problem finding / identifying Noticing or reframing a difficulty in a target already under consideration. This is evidence for generative use of analogy, but not necessarily generation of new target domains from a solution structure.
Solution → problem generation Starting from an explicit solution mechanism/archetype and generating an unmet target problem in another domain for which that mechanism might work. This review reserves the phrase for this direction. Generating a source analog, a benchmark question, or a solution to a given problem does not qualify.
Systematic matrix A declared crossing of solution structures with a declared set of target domains, including failures and exclusions rather than only selected successes. Producing many stochastic samples from one prompt is not a matrix unless the crossing and accounting rules are explicit.

The central task-direction taxonomy is:

  1. Source + target supplied → mapping or answer. SME, visual analogy benchmarks, and many LLM evaluations predominantly test this.
  2. Target problem supplied → retrieve/generate source → adapt solution. Classical analogical problem solving, CBR, analogy-mining systems, analogical prompting, and the 2026 scientific-solution pipeline predominantly test this.
  3. Examples supplied → induce schema. Gick–Holyoak comparison studies, SEQL/SAGE/ALIGN, and parts of LISA/DORA test this.
  4. Source/problem supplied → generate an analogous source or thought experiment. Clement’s expert protocols and some language-generation work test constructive analogy generation.
  5. Solution structure supplied → generate target problems. This is the inverse direction relevant to the EoA claim. It is not interchangeable with the four tasks above.

Synthesis

1. Structural mapping explains what should be preserved

Structure-mapping theory treats analogy as alignment of relational systems rather than a count of shared object attributes. Its most important design commitments are structural consistency, role-based correspondence, and systematicity: connected systems of higher-order relations support stronger mappings and richer inferences than isolated matches. SME operationalizes these commitments by generating local match hypotheses, assembling globally consistent mappings, scoring them structurally, and producing candidate inferences.

For an EoA-style representation, this argues for an explicit role graph rather than only prose or embeddings. A solution archetype should minimally identify:

  • entities or roles;
  • causal, functional, spatial, temporal, or informational relations;
  • enabling conditions and constraints;
  • the intervention or transformation;
  • the resulting state and value criterion;
  • known failure modes and boundary conditions.

The mapping score should reward connected causal/functional systems and penalize role conflicts. But SME also clarifies what this does not solve: it compares already represented base and target descriptions. Even later scaling work notes the burdens of large descriptions, learned versus hand-built representations, and integration with other cognitive processes (Forbus et al., 2017). Structural matching is therefore a middle-stage mechanism, not a complete innovation system.

2. Deep retrieval is harder than deep evaluation

The classic empirical result is a retrieval–evaluation dissociation. In Gick and Holyoak’s (1980) convergence experiments, a distant source story often improved solution of the radiation problem when participants were explicitly hinted to use it; spontaneous access was much less reliable. Gentner, Rattermann, and Forbus (1993) showed that surface similarity strongly influenced retrieval while relational structure strongly influenced judged inferential soundness; recognition controls argued that the weak relational retrieval was not simple forgetting.

MAC/FAC turns this into a computational architecture. A cheap, nonstructural content-vector stage (MAC) filters many cases; a costly SME stage (FAC) performs structural comparison on a handful. The architecture explains why purely relational remindings can occur yet remain rare. It also exposes a systems trade-off: exhaustive structural matching over an open literature is expensive, but semantic filtering can discard precisely the remote matches of interest.

Recent benchmarks show that scale alone has not eliminated this bottleneck. AnaloBench, with 340 human-written analogies and 15 tested language models, reports limited gains from model scale when scenarios are long or when the relevant scenario must be recalled from a large pool. ANALOGYKB shows that a million-scale analogy knowledge base can improve recognition, generation, explanation, and out-of-domain tasks, but its demonstrations still retrieve examples for supplied analogy tasks rather than discover new target problems.

Deep retrieval is difficult for at least five separable reasons:

  1. Cue mismatch: memory and vector retrieval are indexed by available wording, entities, and context, while the useful commonality may be a latent causal role pattern.
  2. Representation acquisition: structural matchers require relations and bindings that may not be explicit in prose, images, or data.
  3. Combinatorics: a corpus can contain many objects, relations, granularities, and potential role assignments; the candidate set grows before structural scoring begins.
  4. Goal dependence: the same source–target pair supports different useful mappings under different goals and constraints.
  5. Adaptation uncertainty: a structurally appealing source can rely on omitted mechanisms or boundary conditions, so retrieval cannot be ranked only by elegance of correspondence.

An EoA retrieval pipeline should therefore be hybrid: broad semantic/domain diversification; relational role extraction; structural reranking; and explicit uncertainty and coverage accounting. It should also audit which candidates were eliminated at each stage, because a high final score cannot reveal retrieval blind spots.

3. Encoding and comparison can construct the abstractions retrieval needs

Gick and Holyoak (1983) found that attempts to improve transfer from one source using a summary, verbal rule, or diagram did not produce notable gains. With two analogs, however, participants often induced a schema while describing their similarities; schema quality strongly predicted transfer, and verbal/diagrammatic support became useful. Gentner, Loewenstein, and Thompson (2003) likewise found that comparing cases improved transfer in negotiation learning relative to studying cases separately.

Computational models make related proposals. SEQL progressively aligns new exemplars with stored generalizations and retains unmatched examples. LISA supplies an integrated account of access, mapping, inference, and schema induction using distributed semantic features dynamically bound into propositions (Hummel & Holyoak, 1997). DORA learns explicit predicate-like relational representations from initially nonrelational inputs (Doumas, Hummel, & Sandhofer, 2008); a later model learns relational content and format without supervision and exhibits zero-shot cross-domain generalization in controlled tasks (Doumas et al., 2022). ALIGN combines probabilistic analogical generalization with automatically retrieved near misses, improving concept learning on sketch and geospatial datasets.

These results support a contrastive curriculum for solution archetypes:

  • two or more positive instantiations from varied domains;
  • a surface-similar but structurally wrong hard negative;
  • a structurally similar but semantically distant positive;
  • a near miss that violates one enabling condition;
  • explicit relational labels and role bindings;
  • a repaired case showing which modification restored validity.

This curriculum should help separate invariant mechanism from incidental implementation. It also guards against “archetypes” that are merely polished summaries of one case.

4. Retrieval and mapping are not enough: adaptation is a second failure point

CBR supplies a mature decomposition of experience reuse. Aamodt and Plaza (1994) summarize the cycle as retrieving a case for a new problem, reusing its information, revising and testing the proposed solution, and retaining the resulting experience. CBR thus contains operations that matter for EoA—solution criticism, feedback, failure memory, and iterative learning—but it normally begins with a described problem and searches mostly domain-specific case libraries.

Derivational analogy in PRODIGY transfers more than a final plan. It stores annotated problem-solving traces, including justifications, abandoned alternatives, and failures, then tests those justifications during replay, adapts when they no longer hold, and falls back to general problem solving when necessary. Demonstrations cover transportation, process-job planning, and extended STRIPS settings. This is a strong precedent for preserving repair trajectories and decision rationales. Its direction remains a supplied initial state and goal → retrieved derivation → adapted plan.

Proteus integrates retrieval, mapping, transfer, storage, and dynamically renewed mappings when transferred procedure steps create new objects. It also illustrates a boundary: visuospatial representations can transfer procedures in some tasks, but causal and functional knowledge is needed to evaluate many transfers; its evaluation subtask was not fully automated through visuospatial reasoning alone.

Behavioral evidence similarly warns that a successful mapping does not finish the work. In mathematical analogical problem solving, Novick and Holyoak (1991) distinguished mapping from procedure adaptation: successful mapping was insufficient for successful transfer, and adaptation was a major source of difficulty. Novick (1988) predicted and observed expertise-sensitive positive and negative transfer: experts are more likely to benefit from structural similarity without surface overlap, while novices are more vulnerable when surface overlap masks structural mismatch.

For EoA, every projected solution should therefore include:

  • which source relations were preserved;
  • which relations were dropped or added;
  • target-specific constraints that force adaptation;
  • a causal explanation of why the adapted mechanism should still work;
  • a failure test that would distinguish real transfer from metaphorical resemblance;
  • a record of rejected mappings and attempted repairs.

5. Far analogy is possible, but distance does not guarantee value

Laboratory studies emphasize spontaneous retrieval failures; naturalistic studies complicate any claim that deep analogies are generally inaccessible. Clement (1988) observed experts generating analogies through association, principles, and especially transformations of the original problem; some analogs were newly invented thought experiments rather than retrieved cases. Christensen and Schunn (2007) observed frequent within- and between-domain analogies in engineering design. Analogies served to identify problems, solve them, and explain concepts; problem-identifying analogies were mainly within-domain, explanatory analogies mainly between-domain, and solution analogies mixed. Goldwater et al. (2021) found that science expertise and cross-disciplinary training increased generation and sorting by causal structure.

These findings support three qualifications:

  1. expertise, active goals, and rich domain knowledge can create retrieval cues missing in simplified experiments;
  2. analogy generation can be constructive, not only retrieval from a fixed case store;
  3. direct evidence for analogy-assisted problem identification exists, but it does not yet show an automated, systematic solution-archetype → new-domain-problem matrix.

Distance is also not a quality metric. A remote analogy may be surprising but causally thin; a nearer analogy may preserve constraints needed for implementation. The 2026 scientific-solution preprint reports a novelty–applicability tension in its judged analogy metrics, reinforcing the need to score distance, structural depth, feasibility, and evidential support separately (Shen et al., 2026).

6. Neural and neuro-symbolic work reduces hand-coding but does not erase task boundaries

End-to-end neural relation modules show that useful relational processing can emerge without classical symbolic rule systems. Santoro et al. (2017) demonstrated a relation-network module on visual question answering, text reasoning, and physical systems. Yet the systems were trained for supplied tasks and questions; the demonstrations are not case retrieval or target-problem discovery.

Abstract-reasoning benchmarks reveal both progress and fragility. In the Procedurally Generated Matrices benchmark, ordinary ResNets generalized poorly, a relational architecture performed better but unevenly across held-out regimes, and auxiliary training to predict symbolic explanations markedly improved generalization (Barrett et al., 2018). Learning Analogies by Contrasting (LABC) showed that carefully constructed alternatives—where only the correct answer shares the source’s relation—can induce cross-domain visual and symbolic analogy in simple networks (Hill et al., 2019). This is strong evidence that data and negative construction can matter at least as much as architectural complexity.

VisiPAM combines pretrained visual representations, attributed graphs, and probabilistic analogical mapping. Without direct training on the part-mapping task it reduced error by 30% relative to a directly trained structured matching model and reproduced qualitative human patterns on cross-category 3D mappings. Both node and edge similarity ablations impaired performance. The result undercuts two simplistic claims: that all structure must be hand-coded, and that object/surface similarity should always be removed. It still receives source and target inputs and transfers part labels; it does not retrieve target problems from open data.

7. LLM evidence is real but contested—and mostly tests supplied problems

Webb, Holyoak, and Lu (2023) reported that GPT-3-family models matched or exceeded human performance on several text-based analogical tasks, including newly constructed matrices and letter-string problems. The result demonstrates substantial behavioral competence on supplied tasks, not a settled human-like mechanism or an innovation pipeline.

Counterfactual tests expose sensitivity to unfamiliar representations. Lewis and Mitchell (2024) found sharp GPT-model declines on variants designed to preserve the abstract rule while changing familiar symbol systems, with human performance remaining high. Webb, Holyoak, and Lu (2025) replied that code-augmented GPT-4 reached roughly human performance on a shuffled-alphabet task because code repaired an auxiliary indexing/counting weakness; without code, performance remained lower. Stevenson et al. (2026) then compared adults, children, and multiple LLM families across Latin, Greek, and novel symbol alphabets: human performance was stable, while every reported LLM degraded from familiar to near- and far-transfer forms. The disagreement is informative rather than resolvable by one score: system scaffolding, representation, auxiliary demands, and contamination controls materially alter the result.

LLM prompting systems also require causal controls. Yasunaga et al. (2024) prompt a model to self-generate relevant exemplars before solving a supplied math, code, or BIG-Bench problem and report gains over zero-shot and manual few-shot chain-of-thought. Qin et al. (2025) find that random self-generated examples can match or outperform relevant examples on some tasks and that exemplar correctness is a key driver. Thus “analogical prompting improved accuracy” does not establish that relational relevance caused the improvement unless the study includes random, surface-matched, structurally matched, and correctness-matched controls.

Finally, open-context retrieval remains hard even when fixed analogies are solvable. AnaloBench’s long-context and large-pool conditions show limited scaling gains, and 2026 far-transfer results show difficulty constructing on-the-fly representations of unfamiliar symbol systems. For EoA, this argues for external structured memory, explicit role extraction, retrieval audits, and executable checks rather than relying on a model’s unobserved parametric recall.

Nearest precedents to the EoA workflow

A. LLM analogical reasoning for scientific solution generation (nearest end-to-end partial precedent)

Shen, Druckmann, and Zou (2026) evaluate three LLMs on 50 biomedical research problems. Their analogical-reasoning pipeline:

  1. receives a target research problem;
  2. extracts objects and relations;
  3. generates cross-domain problem analogies with explicit object mappings and shared relations;
  4. searches the analogous domain for a method and transfers it back as a candidate solution;
  5. checks novelty using Semantic Scholar retrieval, SPECTER reranking, and an LLM judge;
  6. validates judge alignment with human annotations;
  7. implements one proposal in each of four biomedical case studies; and
  8. releases prompts, code, and an analogy dataset.

Across the reported 50-problem evaluation, analogical prompting increased domain and solution diversity and received higher novelty judgments than no-domain and cross-domain baselines; human raters preferred the analogical solution for novelty in 78% of pairwise comparisons and judged proposals reasonable 67% of the time. The four implementations report quantitative gains, including new best results on subsets/metrics in two case studies. These are meaningful system demonstrations, not only vignette ratings.

Limitations relative to EoA are decisive: it starts from the problem, not the solution structure; it samples candidate analogies rather than crossing a declared archetype–domain matrix; prior-art checking evaluates whether the proposed transfer has appeared in retrieved papers but does not establish adoption need or stakeholder usefulness; only four proposals are implemented; and the work is a May 2026 preprint whose own limitations identify feasibility evaluation and execution-grounded end-to-end integration as unfinished. It should be treated as a must-compare contemporary baseline, not as evidence that inverse problem generation is already solved.

B. Scalable analogical innovation with crowds and AI (nearest retrieval/abstraction program)

Kittur et al. (2019) decompose analogical innovation into abstracting a target problem, searching for analogies, and applying them to generate solutions. Their crowd- and AI-supported studies address fixation, scale, and multiple constraints; intermediate problem schemas help find more distant useful inspirations, and learned representations support search over large natural-language collections. Gilon et al. (2018) target search to a supplied design need, and related scientific search work supports ideation through analogical retrieval.

This program is a close precedent for explicit abstraction, distant retrieval, hybrid human–machine roles, and empirical creativity evaluation. Its principal direction remains target need → abstract problem → source inspiration → target solution. It does not make solution archetypes the fixed row variable of a target-problem-generation matrix.

C. SME + MAC/FAC + analogical generalization (nearest modular cognitive architecture)

The SME family supplies a modular stack: SME for mapping/inference, MAC/FAC for retrieval, and SEQL/SAGE/ALIGN for generalization and discrimination. It offers explicit representations, interpretable mappings, candidate inferences, scalable approximations, incremental comparison, and near-miss learning. This is the strongest classical precedent for EoA’s structural core.

Relative limitations are representation acquisition, reliance on a supplied query/target, and lack of an integrated usefulness/prior-art/implementation loop. The stack suggests components and baselines, not an existing inverse-innovation workflow.

D. CBR and derivational analogy (nearest adaptation, criticism, and repair precedent)

CBR’s retrieve–reuse–revise–retain cycle and PRODIGY’s replay of annotated justifications/failures are the closest precedents for proposal repair and longitudinal case memory. Proteus adds dynamic remapping across intermediate procedure states. These systems show that storing only a final answer throws away information needed for robust reuse.

Their cases are generally indexed to solve a supplied problem in a represented domain. They do not perform open-ended cross-domain target-problem generation or external adopter/usefulness discovery.

E. Human problem identification by analogy (nearest reverse-direction behavioral evidence)

Christensen and Schunn’s in-vivo engineering study directly observes analogy serving a problem-identification function. Clement’s protocols show experts constructing new analogical thought experiments through transformation. These findings make the reverse direction cognitively plausible and warn against defining analogy only as memory retrieval.

They are observational human evidence, not reproducible automated system demonstrations. Problem-identifying analogies were mainly within-domain in Christensen and Schunn’s sample, which also cautions against assuming that problem finding will automatically be cross-domain.

Evidence table

Source and status Task direction and representation Demonstrated capability Target-problem generation? Main boundary for this review
Gentner 1983, peer-reviewed theory supplied base + target → mapping; predicate/relational descriptions structural consistency and systematicity account No Theory, not retrieval, adaptation, or a system pipeline
Gick & Holyoak 1980, controlled human experiments target problem + remembered story → solution distant solution transfer, strongly aided by a hint No Shows access bottleneck; small laboratory tasks
Gick & Holyoak 1983, controlled human experiments multiple solved examples → schema → target solution comparison-induced schema predicts transfer No Schema induction depends on encoding conditions
Gentner, Rattermann, & Forbus 1993, controlled human experiments memory probe → retrieval; supplied pair → soundness judgment surface-sensitive access vs relation-sensitive evaluation No Narrative materials; does not prescribe corpus retrieval engineering
Novick 1988, controlled human experiments prior math problem → new math problem expertise-sensitive positive and negative transfer No Within a formal domain; expertise proxy and task representation matter
Clement 1988, expert protocols scientific problem → generated analogy/thought experiment constructive analogy by transformation, association, principle Partial: analogical cases are generated Not an automated system; generated analog is used on the original problem
Christensen & Schunn 2007, in-vivo design study active design target ↔ spontaneous analogies problem identification, solution, and explanation functions Partial: identifies problems in active design Observational; problem-identification analogies mainly within-domain
SME, implemented model structured base + target → mapping/inferences explicit correspondence, scoring, candidate inference No Candidate retrieval and representations external; classic examples small/hand-coded
MAC/FAC, implemented model target/query + represented case library → source retrieval scalable two-stage surface/content filter plus structural reranking No Query target supplied; fixed represented memory; filter can miss deep cases
LISA, implemented cognitive model represented cases/target → access, mapping, inference, learning dynamic binding with distributed semantics; schema induction No Capacity limits and structured propositional inputs
DORA 2008 and Doumas et al. 2022, implemented models nonrelational examples → learned predicates; learned domain → novel domain relation discovery and controlled zero-shot cross-domain generalization No Controlled perceptual/game tasks; no open literature or innovation validation
SEQL, implemented model example stream → relational generalizations progressive structural abstraction and exemplar retention No Representations supplied; category learning, not proposal generation
ALIGN, implemented model positive examples + retrieved near misses → concept discriminative relational concept learning on sketches/maps No Concept classification; domain-specific encoders and datasets
Aamodt & Plaza 1994, framework plus system survey target problem + cases → reuse/revise/retain adaptation, testing, failure memory, incremental learning No Usually intra-domain and target-first; review rather than one experiment
Veloso & Carbonell 1993, implemented planner initial state + goal → retrieved derivation → plan replays justifications/failures, adapts or falls back to search No Symbolic planning domains; solution performance/search reduction, not problem discovery
Proteus 2008, implemented system target visual problem + source cases → transferred procedure retrieval, dynamic remapping, procedure transfer and storage No Target supplied; causal/functional evaluation not automated by visual knowledge alone
Kittur et al. 2019, peer-reviewed human–AI program target problem → schema/search → inspirations/solutions scaled analog search, fixation reduction, crowd/AI decomposition No Perspective integrates several studies; target-first and human-mediated
Barrett et al. 2018 / Hill et al. 2019, implemented neural benchmarks fixed analogy/matrix → answer relational inductive bias, explanation supervision, hard-negative curriculum improve OOD performance No Closed synthetic tasks; no retrieval, adaptation, or problem formulation
VisiPAM 2023, implemented neuro-symbolic model visual source + target → part mapping learned attributed graphs plus zero-shot probabilistic mapping No Source/target supplied; perceptual part transfer only
Webb et al. 2023, LLM experiments prompted fixed analogy → answer strong performance on several novel text analogy tasks No Behavioral competence; mechanism, contamination, and robustness debated
Lewis & Mitchell 2024, Webb et al. 2025, Stevenson et al. 2026 familiar/counterfactual source–target tasks → answer maps robustness boundary and effect of code scaffolding No Conflicting interpretations; small formal tasks; code changes the evaluated system
Yasunaga et al. 2024 and Qin et al. 2025 target problem → self-generated exemplars → answer prompting gains, but random-example controls weaken analogy attribution No Exemplars may aid computation through correctness or context rather than structure
AnaloBench 2024, benchmark long scenario / large pool → recall and analogy judgment directly probes long-context and needle-in-haystack analogy No Recognition/application benchmark; 340 curated analogies
Shen, Druckmann, & Zou 2026, preprint with code/data biomedical target problem → structured distant analogy → transferred solution → novelty search/selected implementation closest integrated LLM analogical-innovation demonstration; 50-problem evaluation and four implementations No: it generates analogous source problems and target solutions Unreviewed; biomedical-only; target-first; limited feasibility/adoption validation; no systematic inverse matrix

Implications for EoA claims and experiments

Claims that are safe if stated narrowly

Established background

  • Relational structure can support transfer across surface-dissimilar cases.
  • Retrieval, mapping, inference, adaptation, and evaluation are separable operations with different failure modes.
  • Deep retrieval is harder than recognizing a deep match once candidates are juxtaposed.
  • Comparison across multiple cases, relational labels, hard negatives, and near misses can improve abstraction and discrimination.
  • Computational systems have implemented each of these component operations in restricted settings.
  • Modern neural/LLM systems can perform substantial analogical work, but performance depends on representation, curriculum, context size, scaffolding, and controls.

Supported interpretation

  • An explicit solution-archetype representation is likely to improve auditability and structural reranking compared with unconstrained prose prompting alone.
  • A hybrid semantic-retrieval → relational-extraction → structural-reranking architecture is better motivated than either embeddings or exhaustive structure matching alone.
  • EoA should store failed mappings, near misses, and repair attempts, not only successful outputs.
  • A training/evaluation curriculum should include distant positives, surface twins, near misses, and counterfactual aliases.

Provisional gap

  • This review did not locate an artifact-bearing, reproducible system that integrates solution-archetype-first target-problem generation, systematic domain crossing, complete proposal adaptation, prior-art and user/adopter scrutiny, falsifiable next steps, and retained repair trajectories.

Open questions

  • Whether an explicit archetype actually improves the validity and usefulness of generated target problems, rather than only diversity or apparent structural depth.
  • Whether a declared matrix discovers better opportunities than stochastic broad prompting at equal compute and review budget.
  • Whether structural reranking adds value after strong language-model or embedding retrieval.
  • Whether prior-art search can reliably distinguish a genuinely unused transfer from vocabulary mismatch or search failure.
  • Whether external adopters judge the generated problems as real, important, and actionable.

Claims to avoid

  • “EoA invents analogical transfer,” “is the first analogy engine,” or “is the first system for cross-domain innovation.”
  • “No prior system generates problems.” Problem-generation systems exist in education, planning, self-play, and data synthesis, and human analogy studies observe problem identification. The narrower question is solution-archetype-first, cross-domain unmet-problem formulation with the rest of the EoA workflow.
  • “LLMs cannot reason analogically” or “LLMs have human-like analogy.” Current evidence supports neither categorical statement.
  • “Surface similarity is irrelevant.” It is a powerful retrieval cue and can be a useful mapping constraint; the risk is allowing it to override causal/relational fit.
  • “Farther is better.” Distance, novelty, structural depth, feasibility, and usefulness must be scored separately.

Required baselines and ablations

For each solution archetype × target-domain cell, compare:

  1. No-archetype LLM: ask directly for important problems and solutions in the target domain.
  2. Cross-domain prompt without explicit structure: request ideas from distant fields without role extraction.
  3. Semantic retrieval only: embedding or keyword search with matched corpus budget.
  4. Explicit structural map: archetype roles/relations plus target role extraction and a declared mapping score.
  5. Random archetype control: same workflow with an irrelevant archetype.
  6. Surface twin: a semantically close but structurally incompatible source.
  7. Near-miss/falsifier control: one enabling relation deliberately violated.
  8. Human expert or crowd hybrid: experts abstract, retrieve, or critique at a defined stage, following the Kittur decomposition.
  9. Problem-first analogical baseline: the Shen et al. direction—given the target problem, generate distant analogous problems and import solutions.
  10. With/without prior-art retrieval and with/without executable checking.

Equalize model, sampling budget, corpus, and critic budget where possible. Report all matrix cells, including invalid, duplicated, already-solved, harmful, trivial, and untestable outputs.

Evaluation measures

The literature suggests decomposing evaluation rather than asking for one “analogy quality” score:

  • retrieval recall: does the candidate set contain known held-out structural analogs?
  • surface robustness: performance on renamed entities, counterfactual symbols, and topic-shifted paraphrases;
  • structural precision: role/edge correspondence accuracy and preservation of higher-order causal systems;
  • hard-negative rejection: rejection of surface-similar structural mismatches;
  • adaptation validity: whether target-specific modifications satisfy constraints;
  • novelty/prior art: blinded literature search with calibrated recall, not only an LLM judgment;
  • problem reality: independent domain experts or potential adopters confirm the problem exists;
  • importance and usefulness: separate ratings for impact, feasibility, and willingness to investigate/adopt;
  • falsifiability: a concrete observation, experiment, prototype, or benchmark that could disconfirm the proposal;
  • repair value: whether critique leads to a measurably better second proposal;
  • coverage and failure accounting: every declared matrix cell and retrieval-stage exclusion is represented;
  • out-of-domain generalization: held-out domains and held-out relation combinations, not only new entities within familiar templates.

Suggested synthetic curriculum

A useful pre-deployment benchmark can be generated from explicit relational graphs while preventing topic leakage:

  1. sample an archetype graph with roles, constraints, intervention, result, and failure edge;
  2. instantiate it in multiple semantically distant vocabularies;
  3. create surface-matched negatives by swapping one causal relation;
  4. create near misses by removing an enabling condition;
  5. hold out whole domains and relation combinations;
  6. add counterfactual aliases/symbols to test on-the-fly representation;
  7. require both directions: problem → solution transfer and solution → problem formulation;
  8. require an explicit map, an adapted proposal, and a falsifier;
  9. evaluate retrieval from a distractor corpus before evaluating mapping; and
  10. include a repair round using critique and failed-case memory.

Success on this curriculum would not establish real-world innovation, but it would diagnose which component fails before expensive field evaluation.

Contradictory and boundary evidence

  1. Laboratory retrieval failure versus naturalistic expert analogy. Gick–Holyoak and Gentner–Rattermann–Forbus show strong surface control over spontaneous retrieval. Clement, Christensen–Schunn, and Goldwater et al. show that experts in active tasks can generate distant or causal analogies relatively often. The best synthesis is conditional: deep retrieval is difficult under sparse cues and weak encoding, but expertise, active goals, transformation, and rich representations can change the cue environment.
  2. Symbolic structure versus learned representations. SME/LISA/DORA emphasize explicit roles and bindings; relation networks, LABC, and visiPAM show that relational behavior can emerge from learned features or hybrid graphs. VisiPAM also shows that entity similarity and relation similarity can jointly help. The evidence favors representational and training constraints, not a categorical symbolic/neural divide.
  3. Strong LLM analogy scores versus counterfactual failures. Webb et al. report human-level or better results on several tasks; Lewis–Mitchell and Stevenson et al. find large drops with unfamiliar alphabets/symbols; Webb et al. 2025 recover one counterfactual task using code. These results disagree about the interpretation of failure, but agree that the evaluated system boundary—including tools—must be specified.
  4. Analogical prompting versus random-context effects. Yasunaga et al. report gains from self-generated relevant exemplars; Qin et al. show random examples can be equally or more useful on some tasks. This directly requires relevance- and correctness-matched controls before crediting structural analogy.
  5. Distance/novelty versus applicability. The Shen et al. system produces more distant and novel proposals, but reports lower applicability scores for more adventurous conditions and acknowledges that feasibility checking remains a bottleneck. A diversity engine is not by itself a validity engine.
  6. Problem identification evidence versus inverse-system evidence. Analogy clearly can help human designers identify problems. The observed function usually occurs within an already active design context and, in Christensen and Schunn’s sample, mainly within-domain. It supports plausibility of inverse generation without demonstrating the proposed automated matrix.
  7. Scalable retrieval versus complete structural search. Kittur et al. show that simplified schemas and hybrid human–AI processes can search large repositories usefully; MAC/FAC and AnaloBench show why deep retrieval remains selective. Scalable approximation is possible, but its missed-analogy rate must be measured.

Limitations of this review

  • The search was systematic by concept and citation network but was not a registered database review. It did not run exhaustive Scopus, Web of Science, patent, dissertation, or commercial-system searches.
  • The remit was cognitive science and computational analogy. Design-by-analogy, biomimetics, TRIZ, automated science, recommender systems, and problem-generation literatures overlap and are covered elsewhere in the parent review; only boundary sources needed to assess task direction appear here.
  • Paywalls limited full-text inspection for some classic articles, although primary abstracts, author manuscripts, or institutional copies were used. Central claims were checked against primary papers rather than accepted solely from surveys.
  • Benchmark results are not directly comparable across representations, models, prompts, tools, and contamination controls.
  • The May 2026 scientific-solution system is an unreviewed preprint. Its code/data release improves inspectability but does not substitute for independent replication.
  • “Solution → problem generation” is an operational distinction adopted for this review; adjacent fields may use problem generation, problem finding, need finding, or opportunity discovery differently.
  • No claim of historical priority or uniqueness should be inferred from the provisional gap. Unindexed, proprietary, unpublished, or differently named systems may exist.

Linked bibliography

Foundational behavioral and theoretical work

Classical and hybrid computational models

Neural, neuro-symbolic, benchmark, and LLM work