In 1952, Otto Struve proposed that a planet crossing the disk of a distant star would cause a periodic, predictable dimming of its light. The observation that sparked this idea was unremarkable — stars dim. But Struve connected that ordinary fact to a causal chain that would eventually enable the detection of thousands of exoplanets. No logical rule compelled the leap from "stellar flux decreases" to "a body may be occluding the source." It was, by most accounts, an act of intuition.
What exactly happened in that inferential moment?
The standard accounts fall short
The philosophical vocabulary for describing such moments — "insight," "abduction," "System 1 processing" — is descriptively useful but computationally opaque. Kahneman's dual-process theory correctly captures the phenomenological speed of insight, but offers no account of its content: why does fast processing sometimes produce a correct cross-domain connection rather than a stereotypical response?
Gentner's Structure Mapping Engine (1983) formalizes analogy as the alignment of relational structures across domains — one of the most rigorous accounts of cross-domain inference we have. But it operates on pre-specified symbolic structures and does not address how those structures are extracted from raw observations.
Pearl's causal hierarchy provides a principled account of causal representation, but the causal inference literature focuses on reasoning within a single domain given a known causal graph, not on recognizing that two graphs from different domains share the same skeleton.
What is missing is a framework that begins with raw observations, extracts causal representations, and performs matching across domains.
Intuition as causal structure matching
We propose that cross-domain intuitive insight consists of three operations performed in rapid, often unconscious succession:
1. Causal abstraction. Given an observation, extract a causal representation that captures its essential structure, stripped of domain-specific surface features. Struve's observation becomes: *"a measurable property of an emitter varies periodically, consistent with a regularly-interposed occluder."*
2. Cross-domain search. Search a repository of causal representations drawn from other domains for structurally isomorphic matches. The occlusion template retrieves: *"when an opaque object periodically interposes between a source and observer, the observed flux decreases."*
3. Hypothesis generation. When a match is found, generate a hypothesis: the mechanisms underlying the two domains may be related. A planetary body may be occluding the star.
This account is explicitly causal rather than merely associational. The connection is not that "star dimming" and "shadow" co-occur in text corpora, but that they share an underlying causal skeleton. This is why the insight is generative — it licenses predictions and interventions, not just pattern completion.
It also makes the role of domain expertise precise. Causal abstraction requires understanding a domain well enough to extract its causal structure. A physicist and a painter looking at a dimming star will extract different representations and therefore retrieve different matches. Intuition is not domain-free pattern recognition; it is structure-sensitive inference grounded in expert knowledge.
The one-to-many problem
The formalization immediately reveals a central computational challenge. A single observation is typically consistent with multiple causal representations. A dimming star could be caused by an orbiting planet, a stellar companion in eclipse, an accretion disk fluctuation, interstellar dust, instrumental noise, or intrinsic stellar variability. Each generates a distinct causal graph.
If one observation produces $m$ candidate causal representations and the repository contains $n$ representations each with $k$ abstraction levels, the exhaustive matching space has size $O(m \cdot n \cdot k^2)$. For realistic values — $m = 10$, $n = 10^4$, $k = 5$ — this is computationally prohibitive.
This is not merely a practical inconvenience. It is philosophically informative. The intractability of brute-force matching explains why genuine cross-domain insight is rare, domain-sensitive, and expertise-dependent. The requisite information is available; the search space is simply too large to traverse without powerful constraints.
What makes human intuition tractable
Two mechanisms collapse the search space to manageable size:
Hierarchical abstraction converts matching from a single exhaustive search into a cascade of increasingly specific filters. At the coarsest level, "a property of an emitter changes over time" matches a vast number of causal representations — cheap but uninformative. At finer levels, *"the change is periodic" and then "consistent with a geometric occlusion model"* progressively narrow the candidates. Complexity drops from $O(n \cdot k^2)$ to $O(n \cdot k \cdot \log k)$, with early termination for mismatches.
Problem-directed pruning restricts the search to representations relevant to the agent's current goal. An instrument engineer observing a dimming star searches causal representations involving electronic noise and calibration errors. Struve, whose active problem was the population of bodies orbiting other stars, searches astrophysical representations involving orbital mechanics. The problem state $Q$ restricts matching to a subspace $\Gamma_Q \ll \Gamma$.
Newton's apple produced an insight about gravity not because Newton was uniquely perceptive, but because his active problem — accounting for the Moon's orbit — had already primed a specific region of his causal representation space. The problem itself was the search algorithm.
An important implication follows: *intuition is not a capacity that operates independently of deliberate reasoning.* The slow, effortful work of formulating a problem — defining what is surprising, what requires explanation, what the relevant variables are — is precisely what makes fast intuitive matching tractable. The two processes are complements, not alternatives.
Why current AI systems cannot do this
Large language models can describe the history of transit photometry, explain the physics of stellar occultation, and discuss the statistical methods used in exoplanet detection. What they cannot do — at least not robustly — is perform the original act of connection: given only the observation of periodic stellar dimming and a repository of causal knowledge from other domains, generate the hypothesis that a planetary body may be responsible.
The reason is architectural. LLMs learn distributional associations over linguistic tokens — text that describes causal relationships — without learning causal representations in the sense required for structure matching. This is the distinction Pearl draws between the associational and interventional levels of the causal hierarchy.
Embedding similarity, while capturing something real, is insufficient. Two concepts may be statistically associated without sharing causal structure (correlation without causation), and may share causal structure without being statistically associated if they appear in different textual contexts. Transit photometry and everyday shadow-casting share causal structure but would not, in 1952, have appeared in proximity in any text corpus.
What would it take?
Building systems capable of genuine cross-domain insight would require, at minimum:
- A mechanism for extracting causal representations from observations - A cross-domain repository of such representations - A matching algorithm that can identify structural isomorphisms - A problem-representation module that constrains the search
None of these components individually is beyond the current state of the art. Causal representation learning (Schölkopf et al. 2021) addresses the first. Structure mapping theory (Gentner 1983) addresses the third. Library learning systems like DreamCoder (Ellis et al. 2021) address hierarchical abstraction within single domains. The challenge is their integration — and, critically, extending them to operate across domain boundaries.
The cross-domain matching of learned causal representations is the computational core of what we experience as scientific intuition. It remains largely unaddressed as a unified research problem. Framing it precisely — as we have attempted here — is a first step toward building systems that do not merely describe past discoveries, but generate new ones.
---
*This post summarizes a longer manuscript: "Intuition as Causal Structure Matching: A Computational Account of Cross-Domain Insight."*