Transformational Creativity in Science: A Graphical Theory

Jonah Black; Lav R. Varshney; Samuel Schapiro

REVIEW 3 major objections 3 minor 26 references

Transformational Creativity in Science: A Graphical Theory

T0 review · 3 major / 3 minor · reviewed 2026-08-16 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Modifying an axiom changes at least as many downstream constraints as modifying any rule, so axiom edits have the greatest transformative potential in a scientific conceptual space.

desk verdict A clean but near-tautological DAG formalization of transformational creativity—axioms as sinks make the main theorem an artifact of the arrow convention, yet the paper is an honest, usable scaffold for AI-discovery discussions. read the letter →

arxiv 2504.18687 v2 pith:ALTISH6Q submitted 2025-04-25 cs.AI

Samuel Schapiro , Jonah Black , Lav R. Varshney This is my paper · ORCID

classification cs.AI

keywords transformationalcreativityconceptualspacedirectedacyclicgraphaxiommodificationparadigmshiftscientificrevolutiontransformationAIdiscovery

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper formalizes the intuition that the most transformative changes in science are changes to a field's foundations. It represents a scientific paradigm as a directed acyclic graph (a dependency network with no cycles) whose vertices are constraints on possible artifacts, with axioms as the foundational sink nodes and rules as nodes that depend on earlier constraints. The central result is that the node with the most dependents—and thus the highest 'transformative potential' under modification—is always an axiom, because every rule has a prerequisite that inherits all of its dependents and then some. This gives a graph-theoretic reason why paradigm shifts such as replacing absolute space and time with spacetime are so consequential, and it points AI discovery systems toward interventions on axioms rather than derived rules.

What carries the argument

The central object is a scientific conceptual space $S=(V,E)$, a finite directed acyclic graph with $V\subseteq \mathcal{P}(L)$ for a formal language $L$, where each vertex is a constraint subset of $L$ and an edge $(u,v)$ exists exactly when $u\subseteq v$, so $v$ is a necessary condition for $u$. Axioms are the sink nodes, the vertices with no outgoing edges; rules are all other vertices. The quantity doing the work is $T_p^{\mathrm{mod}}(v)=|\mathrm{depends}(v)|$, where $\mathrm{depends}(v)$ is the set of vertices $u$ for which a directed path runs from $u$ to $v$—the constraints that cannot be invoked without first assuming $v$. Because the graph is acyclic and every non-axiom has at least one prerequisite, the dependent set of any rule is strictly contained in the dependent set of that prerequisite, pushing the maximum of $T_p^{\mathrm{mod}}$ to an axiom. This identity converts the informal slogan 'change the enabling constraints' into a theorem.

What would settle it

Reconstruct the DAG of a well-documented field from its textbooks or knowledge graph, compute the dependent count for every vertex, and locate the change behind a historical revolution. If a revolution was produced by modifying a vertex that is not an axiom and that has fewer dependents than some unchanged axiom, the dependent-count measure fails as a proxy for transformative impact. The theorem itself is mathematically true inside the model; this observation would falsify the model's claim about real science.

Watch

Extended reading notes

Core claim

The paper's central claim is Theorem 4: in a conceptual space $S$ with at least one rule, the vertex maximizing the transformative potential $T_p^{\mathrm{mod}}(v)=|\mathrm{depends}(v)|$ must be an axiom. The proof is a direct consequence of the dependency structure. If a rule $R$ is not an axiom, it has an outgoing edge to some prerequisite $u$; every vertex that depends on $R$ also depends on $u$, while $u$ has at least one additional dependent, namely $R$ itself. Hence $|\mathrm{depends}(u)| > |\mathrm{depends}(R)|$, so no rule can have strictly more dependents than an axiom upstream of it, and a maximum must lie among the axioms. The paper reads this as a formal counterpart to the historical pattern that revolutions restructure foundational constraints rather than merely adjusting derived rules.

Load-bearing premise

The load-bearing premise is that a real scientific paradigm can be represented faithfully as a finite dependency graph with no cycles, and that the number of downstream constraints affected by a change is the right measure of how transformative that change is.

Editorial extensions

If this is right

In any scientific field that admits the paper's DAG representation, the highest-impact point of intervention is an axiom; modifying a derived rule cannot reach more constraints than modifying its prerequisite.
The historical examples in the paper—geocentrism to heliocentrism, Newtonian to relativistic mechanics, and Euclidean to non-Euclidean geometry—can be represented as axiom substitutions $A\to A'$ that cascade through dependent rules.
The framework supports all five standard conceptual-space operations (locating artifacts, rating similarity, inducing a space, generating artifacts, and revising the space), so it gives a computational handle on creative revision.
An AI discovery system built from the paper's pipeline would construct a dependency graph from a field's literature, locate open problems, and use an LLM idea generator to propose axiom modifications, making the intended transformation explicit and inspectable.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A natural extension the paper only sketches: the dependent count measures how many constraints change, not how much each changes. Weighting edges by distance or semantic distance could rank a rule change above an axiom change in some cases, so the theorem is about the chosen measure, not about all notions of transformative impact.
The model is most at home in formal sciences where constraints are genuinely subsets of a formal language; applying it to empirical fields requires an additional step—building the DAG from texts, citations, or expert knowledge—that the paper does not validate.
A testable prediction follows from the framework: landmark scientific transformations should align with modifications of high-dependency, sink-like nodes in a faithful graph reconstructed from historical documents, and this can be checked with existing knowledge-graph or citation data.
The theorem suggests a concrete search heuristic for AI scientists: when anomalies accumulate, mutate axioms rather than tuning mid-level rules; even an approximate DAG might usefully prioritize where to look first.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The short version: this is a DAG formalization of Boden/Kuhn where axioms are sinks, rules are non-sinks, and 'transformative potential' is the number of nodes that have a directed path to a given node. The main theorem—the node with maximum transformative potential is an axiom—is true, but it is true for any finite DAG once you adopt this arrow convention. It is a restatement of the elementary fact that, following outgoing edges from any node, you can always move to a node with strictly more dependents until you hit a sink. The subset-inclusion semantics and the artifact definition play no role in the proof.

What is genuinely nice: the paper is clearly written, it engages Wiggins, Ritchie, Santo et al., and Boden honestly, and it provides a compact vocabulary for talking about 'axiom changes' as the highest-leverage intervention in a dependency graph. The historical illustrations (Copernicus, relativity, non-Euclidean geometry) are apt as illustrations, even if they are constructed post hoc. The AI-super-scientist pipeline is speculative but shows a plausible use of the framework.

Soft spots: first, Definition 1 defines prereq and depends with the same formula, which is an error or at least a serious ambiguity. Second, and more load-bearing, the theorem is an observation about the chosen edge direction. If you draw edges from prerequisite to consequence instead of the other way, axioms would be sources and the maximum-dependents property would not hold. The paper never argues that the sink-as-axiom convention is the right representation of scientific paradigms, nor that |depends(v)| is a faithful measure of historical transformativeness. Those are assumptions, not results. The paper's own conclusion admits the framework may only fit formal domains, which is an honest limitation but also reveals how far the generality claim carries.

Who should read this: people building AI systems for scientific discovery who want a simple formal target for 'change the axioms,' and readers curious about one way to make Boden's notion computational. It is a scaffold, not a measurement or an empirical resolution.

My recommendation: I'd send it to peer review if the venue accepts conceptual contributions, but I would not expect it to survive without heavy revision. The core theorem should be reframed as a design principle, the definitional typo fixed, and the measure defended with independent evidence. As is, it is an idea sketch with a proof that happens to be correct but empty of scientific content.

Referee Report

3 major / 3 minor

Summary. The paper proposes a formal graphical theory of transformational scientific creativity. A scientific conceptual space is modeled as a finite DAG whose vertices are subsets of a formal language and whose edges encode dependency; axioms are defined as sink nodes. Transformative potential of a node is defined as the number of nodes that depend on it (Definition 3). The main formal result, Theorem 4, states that the node with greatest transformative potential must be an axiom. The paper then illustrates the framework with historical examples (Ptolemaic to heliocentric astronomy, Newtonian to relativistic physics, Euclidean to non-Euclidean geometry) and sketches a neuro-symbolic AI pipeline for transformative discovery.

Significance. The paper is clearly written and the proof of Theorem 4 is correct under the authors' definitions. It makes a genuine attempt to connect Boden's and Kuhn's philosophical ideas to a graph-theoretic formalism, and it lists related computational-creativity work. However, the central theorem is a direct and nearly immediate consequence of the modeling choices: axioms are stipulated to be sink nodes and transformative potential is defined as the number of ancestors in the graph. In any finite DAG, a sink must maximize that count. The proof never uses the subset-inclusion condition in Definition 1(2), the artifact definition (Definition 2), or any other scientific content. The historical illustrations are post hoc reconstructions: the DAGs are drawn so that the modified axioms are sinks, so they provide no independent evidence. The paper offers no empirical validation, no testable predictions, and no implementation of the proposed AI pipeline. Its contribution as a theory of scientific creativity is therefore limited to a formal exercise whose scientific relevance remains unestablished.

major comments (3)

[Section 3, Definition 1(4)-(5)] Definition 1(4) and Definition 1(5) define prereq(v) and depends(v) with identical formulas, both as the set of u with a directed path from u to v, yet the surrounding text assigns them opposite meanings (prerequisites versus dependents). The proof of Theorem 4 relies on the graph-theoretic reading of depends(v) as the set of nodes that can reach v (ancestors), but the textual gloss says the opposite. This inconsistency obscures the fact that the theorem follows purely from the edge-direction convention and not from any intended scientific semantics.
[Theorem 4] Theorem 4 is true for any finite DAG once axioms are defined as sinks and transformative potential is the number of nodes that can reach v. The proof does not use the condition (u,v) in E iff u is a subset of v, nor Definition 2, nor any property of scientific rules. Consequently, the theorem is a tautology of the modeling choices: it reflects the definitions of axiom and transformative potential rather than providing independent support for the claim that modifying axioms is most transformative in actual science. The paper gives no argument that real scientific paradigms have the required DAG structure or that ancestor count is a faithful measure of transformative impact, so the central claim about science is unsupported.
[Section 6, Historical Illustrations] The historical illustrations are post hoc narratives rather than formal reconstructions. For example, the Ptolemaic system is represented with 'Earth stationary and central' as a sink node, but nothing in the historical record forces that specific graph; the DAG is chosen so that the modified axiom is a sink. The same applies to the Newtonian and Euclidean examples. Because the graphs are constructed to satisfy the definition, the illustrations are consistent with the model by construction and do not constitute empirical evidence for the theory.

minor comments (3)

[Title] The title on the first page reads 'Graphical Theor y' with an extra space; this should be corrected.
[Definition 1(2)] The edge condition 'u is a further constraint on v, so that v is a necessary condition for u' is confusing: if u is a subset of v, then u is more restrictive and any instance of u is an instance of v, so v is indeed necessary for u. The wording 'further constraint on v' could be clarified to avoid implying that v is a stronger constraint than u.
[Notation] The notation for transformative potential is inconsistent: the text writes 'T p mod(v)', 'Tp_mod(v)', and 'T p mod' in different places; a single notation such as Tp_mod(v) should be used throughout.

Circularity Check

1 steps flagged · score 8.0 of 10

Theorem 4 reduces by construction to the definitions: axioms are sinks and transformative potential counts dependents, so the maximum in any DAG is forced to be an axiom.

self definitional [Definition 1(1), Definition 3, Theorem 4 (with proof)]
"1. Axioms. The set of sink nodes A = { A ∈ V | ∄u ∈ V : ( A, u ) ∈ E } are the axioms of S and have no outgoing edges. ... Tp_mod(v) := |depends(v)| ... Theorem 4 (Modifying Axioms has the Most Transformative Potential). Given a conceptual space S for which V \ A ≠ ∅, the node with the greatest transformative potential when modified Tp_mod must be an axiom."

Axiom is defined as a sink node and transformative potential is defined as the number of dependents. In any finite DAG, if a non-sink R has an outgoing edge (R,u), every dependent of R is also a dependent of u, and R itself is an additional dependent of u, so |depends(u)| ≥ |depends(R)| + 1; therefore a maximum is always attained at a sink. Thus Theorem 4 follows immediately from the definitions and elementary DAG structure, not from the conceptual-space semantics: the proof never uses the subset-inclusion condition of Definition 1(2), Definition 2, or any scientific content. The 'result' that axioms are most transformative is built into calling sinks 'axioms' and measuring transformativeness by dependent count.

full rationale

The paper contains no load-bearing self-citation and no fitted-parameter prediction; its circularity is entirely in the central theorem's dependence on its own modeling choices. Definition 1 stipulates that axioms are exactly the sink nodes, and Definition 3 stipulates that transformative potential is exactly the number of nodes depending on a given node. With these definitions, Theorem 4 is a graph-theoretic fact true of every finite DAG, independent of whether vertices are subsets of a formal language, whether artifacts are defined as triples, or whether the setting has anything to do with science. The proof itself confirms this: it only uses the presence of an outgoing edge from a non-axiom and the path-based notion of dependence. The historical illustrations are post hoc constructions whose DAGs place axioms at sinks, so they illustrate the definitions rather than test them. There is also a smaller definitional slip in Definition 1(4)-(5), where prereq(v) and depends(v) are written with the same formula; even after correcting the apparent typo, the theorem remains a direct consequence of the chosen terminology. Because the graph lemma itself is valid but the scientific conclusion is forced by definition rather than by evidence, the appropriate circularity score is high, though not maximal: the theorem is not strictly identical to a single definition, but its content is fully determined by the definitions and the universal structure of finite DAGs.

Assumptions & free parameters 0 free parameters · 5 assumptions · 0 invented entities

No free parameters or invented empirical entities. The theory's load rests on the modeling assumptions that scientific conceptual spaces are subset-inclusion DAGs and that dependent-node count measures transformativeness.

assumptions (5)

domain assumption Scientific conceptual spaces can be modeled as finite directed acyclic graphs over subsets of a formal language L.
Definition 1 assumes without empirical support that a field's enabling constraints form a DAG.
domain assumption For any two vertices u and v, (u,v) is an edge iff u is a subset of v, encoding dependency as subset inclusion.
Definition 1, property 2; restricts rules and axioms to a pure subset semantics.
domain assumption Axioms are exactly the sink nodes, i.e., vertices with no outgoing edges.
Definition 1, property 1; identifies foundational principles with maximal elements under inclusion.
domain assumption The transformative potential of a node v is measured by the number of nodes that depend on it, Tp_mod(v) = |depends(v)|.
Definition 3; the measure is chosen by the authors and is not derived from data.
standard math Standard finite DAG facts, such as every path being finite and every non-sink having an outgoing edge, hold.
Used in the proof of Theorem 4.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Transformational Creativity in Science: A Graphical Theory." pith.science (2026). https://pith.science/paper/ALTISH6Q

@misc{pith2026250418687,
  author       = {Pith},
  title        = {Pith review of: Transformational Creativity in Science: A Graphical Theory},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ALTISH6Q}},
  note         = {Machine review of arXiv:2504.18687}
}

read the original abstract

Creative processes are typically divided into three types: combinatorial, exploratory, and transformational. Here, we provide a graphical theory of transformational scientific creativity, synthesizing Boden's insight that transformational creativity arises from changes in the "enabling constraints" of a conceptual space and Kuhn's structure of scientific revolutions as resulting from paradigm shifts. We prove that modifications made to axioms of our graphical model have the most transformative potential and then illustrate how several historical instances of transformational creativity can be captured by our framework.

Figures

Figures reproduced from arXiv: 2504.18687 by the authors.

**Figure 1.** Conceptual-space DAGs before (S) and after (S ′ ) the Copernican and Keplerian revolution. All non-axiom artifacts reach at least one axiom (sink node), satisfying the formal definition of a conceptual space [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

26 extracted references · 21 canonical work pages

[4]

Journal of Computational Biology 7:601–620

Using Bayesian networks to analyze expression data. Journal of Computational Biology 7:601–620

work page
[10]

Kon, P . T. J.; Liu, J.; Ding, Q.; Qiu, Y .; Y ang, Z.; Huang, Y .; Srinivasa, J.; Lee, M.; Chowdhury, M.; and Chen, A. 2025. Curie: Toward rigorous and automated sci- entiﬁc experimentation with AI agents. arXiv:2502.16069

arXiv 2025
[12]

T.; Foerster, J.; Clune, J.; and Ha, D

Lu, C.; Lu, C.; Lange, R. T.; Foerster, J.; Clune, J.; and Ha, D. 2024. The AI scientist: Towards fully automated open-ended scientiﬁc discovery. arXiv:2408.06292

arXiv 2024
[15]

Towards a Formal Creativity Theory: Preliminary results in Novelty and Transformativeness

Santo, L. E.; Wiggins, G.; and Cardoso, A. 2024. Towards a formal creativity theory: Preliminary results in novelty and transformativeness. arXiv 2405.02148

work page Pith review arXiv 2024
[17]

G.; and Evans, J

Shi, F.; Foster, J. G.; and Evans, J. A. 2015. Weaving the fabric of science: Dynamic network models of science’s unfolding structure. Social Networks 43:73–85

work page 2015
[18]

Si, C.; Y ang, D.; and Hashimoto, T. 2024. Can LLMs generate novel research ideas? a large-scale human study with 100+ NLP researchers. arXiv:2409.04109

arXiv 2024
[21]

Su, H.; Chen, R.; Tang, S.; Zheng, X.; Li, J.; Yin, Z.; Ouyang, W .; and Dong, N. 2024. Two heads are better than one: A multi-agent system has the potential to improve scientiﬁc idea generation. arXiv:2410.09403

arXiv 2024
[25]

Wang, Q.; Downey, D.; Ji, H.; and Hope, T. 2024. SciMON: Scientiﬁc inspiration machines optimized for nov- elty. In Proceedings of the 62nd Annual Meeting of the As- sociation for Computational Linguistics , 279–299

work page 2024

Show all 26 references

[26]

Wiggins, G. 2006. A preliminary framework for description, analysis and comparison of creative systems. Knowledge-Based Systems 19:449–458

2006
[1901]

Jaccard, P . 1901. ´Etude comparative de la distribution ﬂorale dans une portion des Alpes et des Jura. Bull. Soc. V audoise Sci. Nat.37:547–579

1901
[1962]

Kuhn, T. S. 1962. The Structure of Scientiﬁc Revolu- tions. University of Chicago Press

1962
[1964]

Koestler, A. 1964. The Act of Creation . Macmillan

1964
[1992]

Boden, M. 1992. The Creative Mind. London: Aba- cus

1992
[2000]

Friedman, N.; Linial, M.; Nachman, I.; and Pe’er, D
[2004]

Simonton, D. K. 2004. Creativity in Science: Chance, Logic, Genius, and Zeitgeist . Cambridge University Press

2004
[2006]

Ritchie, G. 2006. The transformational creativity h y- pothesis. New Generation Computing 24:241–266

2006
[2007]

Haven, K. 2007. 100 Greatest Science Discoveries of All Time. Bloomsbury Publishing USA

2007
[2012]

Thagard, P . 2012. Creative combination of represen- tations: Scientiﬁc discovery and technological invention . In Psychology of Science: Implicit and Explicit Processes . Ox- ford University Press

2012
[2014]

Rothenberg, A. 2014. Flight from W onder: An Inves- tigation of Scientiﬁc Creativity . Oxford University Press

2014
[2015]

Aguiar, D.; At˜ a, P .; and Queiroz, J. 2015. Intersemi - otic translation and transformational creativity. Punctum. International Journal of Semiotics 1:11–21

2015
[2018]

Toivanen, J.; J¨ arvisalo, M.; Alm, O.; V entura, D.; V ainio, M.; and Toivonen, H. 2018. Towards transforma- tional creation of novel songs. Connection Science 31:1–29

2018
[2020]

V arshney, L. R. 2020. Limits theorems for creativ- ity with intentionality. In Proceedings of the International Conference on Computational Creativity, 390–393

2020
[2021]

Simonton, D. K. 2021. Scientiﬁc creativity: Discov- ery and invention as combinatorial. Frontiers in Psychology 12:721104

2021
[2023]

Shi, F., and Evans, J. 2023. Surprising combina- tions of research contents and contexts are related to impac t and emerge with scientiﬁc outsiders from distant disciplin es. Nature Communications 14

2023
[2024]

Gu, X., and Krenn, M. 2024. Interesting scientiﬁc idea generation using knowledge graphs and LLMs: Evalu- ations with 100 research group leaders. arXiv:2405.17044

2024 arXiv
[2025]

Intology. 2025. Zochi technical report. https://www.intology.ai/blog/zochi-tech-report

2025

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.

[1] [4]

Journal of Computational Biology 7:601–620

Using Bayesian networks to analyze expression data. Journal of Computational Biology 7:601–620

work page

[2] [10]

Kon, P . T. J.; Liu, J.; Ding, Q.; Qiu, Y .; Y ang, Z.; Huang, Y .; Srinivasa, J.; Lee, M.; Chowdhury, M.; and Chen, A. 2025. Curie: Toward rigorous and automated sci- entiﬁc experimentation with AI agents. arXiv:2502.16069

arXiv 2025

[3] [12]

T.; Foerster, J.; Clune, J.; and Ha, D

Lu, C.; Lu, C.; Lange, R. T.; Foerster, J.; Clune, J.; and Ha, D. 2024. The AI scientist: Towards fully automated open-ended scientiﬁc discovery. arXiv:2408.06292

arXiv 2024

[4] [15]

Towards a Formal Creativity Theory: Preliminary results in Novelty and Transformativeness

Santo, L. E.; Wiggins, G.; and Cardoso, A. 2024. Towards a formal creativity theory: Preliminary results in novelty and transformativeness. arXiv 2405.02148

work page Pith review arXiv 2024

[5] [17]

G.; and Evans, J

Shi, F.; Foster, J. G.; and Evans, J. A. 2015. Weaving the fabric of science: Dynamic network models of science’s unfolding structure. Social Networks 43:73–85

work page 2015

[6] [18]

Si, C.; Y ang, D.; and Hashimoto, T. 2024. Can LLMs generate novel research ideas? a large-scale human study with 100+ NLP researchers. arXiv:2409.04109

arXiv 2024

[7] [21]

Su, H.; Chen, R.; Tang, S.; Zheng, X.; Li, J.; Yin, Z.; Ouyang, W .; and Dong, N. 2024. Two heads are better than one: A multi-agent system has the potential to improve scientiﬁc idea generation. arXiv:2410.09403

arXiv 2024

[8] [25]

Wang, Q.; Downey, D.; Ji, H.; and Hope, T. 2024. SciMON: Scientiﬁc inspiration machines optimized for nov- elty. In Proceedings of the 62nd Annual Meeting of the As- sociation for Computational Linguistics , 279–299

work page 2024