Autonomous Synthesis of Analogous Target Domains

17 concepts7 papers1 proposal

For decades, computational models of analogy operated under a fundamental assumption: both the source domain and the target domain already exist. In this traditional paradigm, the challenge of analogy-making is framed entirely as a retrieval and mapping problem. Given a base concept (e.g., the structure of an atom), a system searches a pre-existing database to find a matching target domain (e.g., the solar system) and aligns their relational structures.

However, this approach suffers from a severe bottleneck. If a system's database does not already contain a suitable target domain, it cannot form an analogy. It cannot innovate.

A growing body of research flips this paradigm on its head, shifting analogy-making from a retrieval task to a generative one. Instead of searching through existing datasets, these systems autonomously synthesize entirely new target domains, concepts, or solutions from scratch. By treating analogy as a generative optimization problem, these frameworks enable true computational creativity and open-ended problem-solving.

The Core Mechanism: Optimizing for Isomorphism

To understand why autonomous target domain synthesis works, we must look at how knowledge can be structurally manipulated. If you strip away the specific surface attributes of a concept—such as the fact that an electron is small and negative—you are left with a structural blueprint of relations (e.g., $A$ orbits $B$; $B$ attracts $A$).

To generate a brand-new analogy, a system must construct a novel network of concepts that satisfies two competing constraints:

  1. Structural Isomorphism: The new network must mirror the relational patterns of the base concept.
  2. Semantic Meaningfulness: The generated entities and relations must make sense in the real world, avoiding random or incoherent combinations.

As demonstrated by Santiago Ontanon et al. in their work on the Automated Generation of Cross-Domain Analogies via Evolutionary Computation, this can be achieved using evolutionary algorithms. In this framework, the "genome" is a semantic network representing a conceptual domain. The system initializes a population of random or sparse networks and evolves them over generations using mutations (adding, deleting, or changing nodes and relations).

[Base Domain] ---> Extract Relational Blueprint (Isomorphism Constraint)
                                   |
                                   v
[Random Seed] ---> Evolutionary Mutations ---> [Synthesized Target Domain]
                                   ^
                                   |
                     Commonsense Knowledge Base (Validity Constraint)

To ensure the results are both structurally sound and meaningful, the fitness function evaluates how well the evolving network maps to the base domain, while external commonsense knowledge bases like WordNet or ConceptNet act as constraints. As Atilim Günes Baydin et al. show in A semantic network-based evolutionary algorithm for computational creativity., employing analogical similarity directly as a fitness measure allows systems to spontaneously invent creative parallels. Rather than just finding relationships within existing structures, the algorithm constructs novel "memes"—structured pieces of information—such as autonomously inventing a scenario that maps the physics of a solar system onto the biological dynamics of a tree and its fruit, as detailed by Santiago Ontanon et al. in A semantic network-based evolutionary algorithm for modeling memetic evolution and creativity.

Direct Generation in Proportional and Linguistic Spaces

This generative philosophy also resolves long-standing limitations in algebraic and linguistic analogy solving, typically framed as completing the proportion $a : b :: c : x$ (read as "$a$ is to $b$ as $c$ is to $x$").

Traditional classifiers or retrieval systems attempt to select $x$ from a predefined vocabulary. However, if the correct answer requires a novel transformation, retrieval fails. In their comprehensive review, Henri Prade et al. emphasize that analogical proportions are uniquely powerful because they can be used to compute and generate entirely novel solutions distinct from any known elements in the dataset, serving as a foundation for creative problem-solving in Analogical Proportions: Why They Are Useful in AI.

Retrieval Approach:   [a : b :: c : ?] ---> Search Vocabulary ---> Select Best Fit (Fails if word is new)
Generative Approach:  [a : b :: c : ?] ---> Character Autoencoder ---> Synthesize New Word Form (Succeeds)

We see a concrete implementation of this in natural language processing. When dealing with morphological analogies—such as predicting the plural or tense transformation of an unseen or made-up word—retrieval-based methods hit a hard wall. To overcome this, Kevin Chan et al. proposed a generative approach in Solving Morphological Analogies Through Generation. By leveraging character-level autoencoders, the model learns the underlying geometric transformations of word structures. Instead of picking a word from a dictionary, it physically synthesizes the characters of the target word form $x$ from scratch. This capacity for analogical innovation allows the system to generate correct linguistic forms for words it has never previously encountered.

Escaping Local Anchoring in Language Models

The necessity of autonomous synthesis becomes even more apparent when moving from simple word forms to high-level conceptual reasoning. Human language is saturated with novel metaphors and fluid conceptual shifts. Early on, John A. Barnden identified that rigid, predefined domain mappings are fundamentally insufficient for true metaphor understanding in An Artificial Intelligence Approach to Metaphor Understanding. Instead, systems must be capable of operating within a "pretence sandbox," dynamically reasoning and generating new analogical frameworks on the fly to interpret original figurative language.

In the era of Large Language Models (LLMs), this generative capability serves a vital role: breaking free from data memorization. LLMs are notorious for "local anchoring"—clinging to highly frequent patterns and literal associations present in their training data, which limits their capacity for original scientific ideation.

To force models out of these local minima, the authors of NOVA-Test: Auditing LLM-Generated Research Hypotheses with Structural Analogy, Prior-Work, and Falsifiability Tests introduced a protocol that mandates cross-domain generative analogy. As David Scott Lewis et al. demonstrate, by forcing the LLM to map underlying mechanisms from completely distant, unrelated fields to a target problem, the framework bypasses standard literature-search prompting. The model cannot simply retrieve an existing solution; it is forced to synthesize a novel conceptual framework, resulting in research hypotheses that are significantly more creative, technically specific, and scientifically valuable.

Why the Generative Approach Works

Ultimately, autonomous target domain synthesis works because it decouples the structure of knowledge from its content. By treating analogy as a blueprint for construction rather than a filter for selection, these systems achieve several key advantages:

  • Overcoming Vocabulary Limitations: Whether generating a brand-new semantic network or a novel word form, the system is not constrained by what is already written in its database.
  • Robustness to Novelty: It can handle out-of-distribution scenarios because it possesses the machinery to build its own context.
  • True Computational Creativity: By optimizing for structural alignment while drawing from diverse commonsense constraints, systems can discover unexpected, cross-domain parallels that human designers did not explicitly encode.

Key Foundational Work

Automated Generation of Cross-Domain Analogies via Evolutionary Computation (2012) — contributed 4 concept elements to this direction.
An Artificial Intelligence Approach to Metaphor Understanding (2007) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_autonomous_target_domain_synthesis_cc19c270,
  title = {autonomous target domain synthesis},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/autonomous-target-domain-synthesis/rdr_cc19c270bd6b4897895613b4f9cfd332},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 1, frozen text):

@misc{lacuna_autonomous_target_domain_synthesis_cc19c270_v1,
  title = {autonomous target domain synthesis},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/autonomous-target-domain-synthesis/rdr_cc19c270bd6b4897895613b4f9cfd332/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-22},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake