arxiv: 2509.18218 · v5 · submitted 2025-09-21 · 💻 cs.AI

Similarity Field Theory: A Mathematical Framework for Intelligence

Kei-Sing Ng This is my paper

Pith reviewed 2026-05-18 14:31 UTC · model grok-4.3

classification 💻 cs.AI

keywords similarity fieldfiber preservationgenerative operatorintelligence definitionsimilarity relationsAI interpretabilitylevel sets

0 comments

The pith

Similarity Field Theory defines intelligence as a generative operator that keeps new entities inside a concept's similarity fiber.

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

The paper posits that similarity relations underpin comprehensible dynamic systems and introduces Similarity Field Theory to formalize them. It models entities with a similarity field S that maps pairs to values in [0,1] and treats concepts as superlevel sets called fibers. Intelligence is defined geometrically: an operator G is intelligent with respect to concept K when it generates new entities that remain in K's fiber. This view reframes both intelligence and interpretability as problems of fiber preservation and composition instead of statistical pattern matching. The paper proves two theorems that limit how similarity fields can evolve without breaking reflexivity or stability.

Core claim

Similarity Field Theory formalizes a similarity field S over a universe of entities, with concepts K inducing fibers as superlevel sets of similarity to K. It defines a generative operator G to be intelligent with respect to K precisely when, given entities already in the fiber of K, G produces new entities that also belong to that fiber. Two theorems follow: asymmetry in S blocks mutual fiber inclusion, and stability of the evolving system implies either an anchor coordinate or asymptotic confinement to the target level within arbitrary tolerance. The framework thereby supplies a geometric language for characterizing intelligent systems.

What carries the argument

The similarity field S: U × U → [0,1] together with fibers F_α(K) = {E ∈ U | S(E,K) ≥ α} as superlevel sets, carried by the generative operator G that must preserve fiber membership.

If this is right

Intelligence and interpretability reduce to geometric tasks of preserving and composing level-set fibers.
System evolution must obey the constraints from asymmetry blocking mutual inclusion and stability requiring anchors or confinement.
AI alignment targets human-observable and human-interpretable versions of concepts such as safety rather than the full underlying concept.
Similarity fields supply a common language for comparing and constructing different intelligent systems.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Mapping embedding spaces of current language models onto explicit similarity fields could expose which concepts their generations actually preserve.
A direct test would define a simple concept such as 'prime number' and check whether a generator stays inside the corresponding fiber.
The approach may connect to existing geometric models of cognition by treating similarity values as distances or angles in an underlying space.
If the framework holds, new metrics for interpretability could be built from the rate at which fibers drift under repeated generation steps.

Load-bearing premise

Transforming similarity relations form the structural basis of comprehensible dynamic systems.

What would settle it

An observed generative process that produces intuitively intelligent outputs yet systematically places new entities outside every fiber of the relevant concept, or a stable similarity-field evolution that remains comprehensible without satisfying either an anchor or confinement condition.

Figures

Figures reproduced from arXiv: 2509.18218 by Kei-Sing Ng.

**Figure 2.** Figure 2: A single neuron as a primitive similarity field. [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗

**Figure 3.** Figure 3: A neural network composes simple similarity fields into a complex one. [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗

**Figure 4.** Figure 4: Bars show Spearman ρ and MAE across 11 categories using a three-model ensemble. Model-class sensitivity and limitations. We also probe two chat-aligned closed-source endpoints under the same A/B scoring protocol (no sampling; deterministic sequence log-probabilities at the same Answer: position). On CSD, the summary is: Model Spearman ρ MAE (pp) gpt-4o-mini 0.770 2.241 gemini-2.0-flash 0.758 2.914 Their ra… view at source ↗

read the original abstract

We posit that transforming similarity relations form the structural basis of comprehensible dynamic systems. This paper introduces Similarity Field Theory, a mathematical framework that formalizes the principles governing similarity values among entities and their evolution. We define: (1) a similarity field $S: U \times U \to [0,1]$ over a universe of entities $U$, satisfying reflexivity $S(E,E)=1$ and treated as a directed relational field (asymmetry and non-transitivity are allowed); (2) the evolution of a system through a sequence $Z_p=(X_p,S^{(p)})$ indexed by $p=0,1,2,\ldots$; (3) concepts $K$ as entities that induce fibers $F_{\alpha}(K)={E\in U \mid S(E,K)\ge \alpha}$, i.e., superlevel sets of the unary map $S_K(E):=S(E,K)$; and (4) a generative operator $G$ that produces new entities. Within this framework, we formalize a generative definition of intelligence: an operator $G$ is intelligent with respect to a concept $K$ if, given a system containing entities belonging to the fiber of $K$, it generates new entities that also belong to that fiber. Similarity Field Theory thus offers a foundational language for characterizing, comparing, and constructing intelligent systems. At a high level, this framework reframes intelligence and interpretability as geometric problems on similarity fields--preserving and composing level-set fibers--rather than statistical ones. We prove two theorems: (i) asymmetry blocks mutual inclusion; and (ii) stability implies either an anchor coordinate or asymptotic confinement to the target level (up to arbitrarily small tolerance). Together, these results constrain similarity-field evolution and motivate an interpretive lens applicable to large language models. AI systems may be aligned less to safety as such than to human-observable and human-interpretable conceptions of safety, which may not fully determine the underlying safety concept.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper defines intelligence as fiber preservation under a generative operator but leaves the similarity field without an update rule when new entities are added.

read the letter

The main thing to know is that this paper sets up Similarity Field Theory with a generative definition of intelligence: an operator G counts as intelligent relative to concept K if it produces new entities that land in the same superlevel set of similarity to K. That is the core move, and it is presented as a geometric reframing away from pure statistics toward fiber preservation. The two theorems follow straightforwardly from the stated reflexivity and allowed asymmetry, so they hold on the given axioms without hidden steps. The setup is internally consistent and gives a clean relational language for talking about concepts as entities that induce fibers. It also sketches a possible link to alignment by suggesting systems should track human-interpretable fibers rather than opaque safety notions. The soft spots are straightforward. The intelligence claim is definitional rather than derived from further properties or tested against data, which limits how much it explains on its own. More importantly, the stress-test concern lands: the framework starts from the idea of transforming similarity relations in dynamic systems, yet supplies no evolution equation or extension rule for S when G adds entities to the universe. Without that, you cannot check whether fibers are actually preserved across steps, so the account stays static despite the premise about evolving systems. This is a clear gap rather than a minor omission. The paper is for readers who already work with relational or level-set models and want to see how they might apply to intelligence or interpretability questions. Someone looking for a formal alternative to statistical views of AI could find the definitions and theorems worth discussing. I would send it to peer review. The formal core is solid enough to merit referee time, and the missing dynamic piece is specific enough that revision could address it directly.

Referee Report

1 major / 2 minor

Summary. The paper introduces Similarity Field Theory as a mathematical framework for intelligence, based on the premise that transforming similarity relations underpin comprehensible dynamic systems. It defines a reflexive directed similarity field S: U × U → [0,1], system evolution via indexed sequences Z_p = (X_p, S^{(p)}), concepts K inducing fibers F_α(K) as superlevel sets {E ∈ U | S(E,K) ≥ α}, and a generative operator G. Intelligence is formalized as G preserving fiber membership when generating new entities. Two theorems are proved: asymmetry blocks mutual inclusion, and stability implies an anchor or asymptotic confinement to the target level. The framework reframes intelligence and interpretability as geometric problems of fiber preservation rather than statistical ones, with suggested implications for AI alignment.

Significance. If the definitions and theorems are sound, the work offers a novel relational-geometric language for characterizing intelligent systems and alignment, emphasizing human-interpretable fibers over raw statistical measures. Credit is due for the parameter-free axiomatic setup and the two direct theorems from reflexivity and asymmetry. However, without empirical tests, predictive derivations, or concrete applications, the significance remains primarily conceptual and foundational rather than immediately actionable for AI research.

major comments (1)

[Definitions of system evolution and generative operator] Definitions (2) and (4) and the evolution Z_p: The manuscript claims to formalize principles governing the evolution of similarity values in dynamic systems but supplies no update rule, extension mechanism, or constraint on S^{(p+1)} when G adds new entities. This omission is load-bearing for the central intelligence claim, as fiber preservation cannot be verified or maintained across steps in a transforming system without it (see also the opening premise on transforming similarity relations).

minor comments (2)

[Theorem statements] Theorem (ii) on stability: The statement is clear but would benefit from an explicit proof sketch or reference to the exact properties used, even if it follows directly from reflexivity and allowed asymmetry.
[Fiber and concept definitions] Notation for fibers: The unary map S_K(E) and superlevel sets are introduced cleanly, but an illustrative example with a small U would improve readability for the geometric reframing.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for their careful reading and constructive feedback on our manuscript. The comment raises a valid point about the level of specificity in our definitions, which we address below by clarifying the intended scope of the framework.

read point-by-point responses

Referee: Definitions (2) and (4) and the evolution Z_p: The manuscript claims to formalize principles governing the evolution of similarity values in dynamic systems but supplies no update rule, extension mechanism, or constraint on S^{(p+1)} when G adds new entities. This omission is load-bearing for the central intelligence claim, as fiber preservation cannot be verified or maintained across steps in a transforming system without it (see also the opening premise on transforming similarity relations).

Authors: We acknowledge that the manuscript presents Similarity Field Theory as an abstract axiomatic framework without specifying a concrete update rule or extension mechanism for S^{(p+1)}. This is by design: the theory is intended as a general relational-geometric language rather than a complete dynamical system with prescribed evolution laws. The sequence Z_p is introduced to index evolving systems, but the precise manner in which new entities generated by G modify the similarity field is left open for instantiation in specific domains (for example, recomputation of similarities in an embedding space or rule-based updates in a symbolic system). The intelligence definition requires only that G produce entities belonging to the relevant fiber with respect to the similarity field at the time of generation; preservation across multiple steps then follows from whatever update rule is chosen in a given application. We will revise the manuscript to add an explicit clarifying paragraph after Definition (4) stating this generality, together with a short illustrative example of a possible update rule in the discussion section on applications to AI. This revision will make the framework's scope and the conditions for multi-step fiber preservation more transparent without altering the core definitions, theorems, or claims. revision: partial

Circularity Check

0 steps flagged

No significant circularity; framework is self-contained via explicit definitions

full rationale

The paper introduces a similarity field S with reflexivity, an indexed evolution sequence Z_p = (X_p, S^{(p)}), concepts K inducing fibers as superlevel sets F_α(K), and a generative operator G, then directly states a generative definition of intelligence as G producing new entities that remain in the fiber of K. Two theorems on asymmetry and stability are presented as derived constraints on the field. No parameters are fitted to data and relabeled as predictions, no self-citations are used to justify uniqueness or load-bearing premises, and no ansatz or known result is smuggled or renamed. The chain proceeds from the opening postulate on similarity relations to the geometric reframing without any definitional loop or reduction by construction. The absence of an explicit update rule for S upon entity addition is a potential gap in the dynamic aspect but does not create circularity in the provided definitions or theorems.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 3 invented entities

The central claims rest on positing a similarity field with reflexivity and directedness, defining concepts as superlevel sets, and introducing a generative operator whose intelligence is defined by fiber preservation. No numerical parameters are fitted to data; the structure is axiomatic and definitional.

axioms (2)

domain assumption S(E, E) = 1 for all E in U (reflexivity)
Stated as part of the similarity field definition in the abstract.
domain assumption Asymmetry and non-transitivity are allowed in the directed relational field
Explicitly noted to permit general relational structures.

invented entities (3)

Similarity field S: U × U → [0,1] no independent evidence
purpose: To serve as the foundational relational structure for entities and their evolution
Newly defined as the core object of the theory.
Concept fiber F_α(K) no independent evidence
purpose: To represent concepts geometrically as superlevel sets
Introduced to formalize concepts within the similarity field.
Generative operator G no independent evidence
purpose: To produce new entities and define intelligence via fiber preservation
Central to the generative definition of intelligence.

pith-pipeline@v0.9.0 · 5886 in / 1599 out tokens · 46121 ms · 2026-05-18T14:31:56.574458+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

an operator G is intelligent with respect to a concept K if, given a system containing entities belonging to the fiber of K, it generates new entities that also belong to that fiber
IndisputableMonolith/Foundation/AbsoluteFloorClosure.lean absolute_floor_iff_bare_distinguishability unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

Theorem 1 (Incompatibility Theorem): asymmetry blocks mutual inclusion

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

56 extracted references · 56 canonical work pages · 6 internal anchors

[1]

Alaska Air Group, Inc. (2025). Alaska Air Group Reports Fourth Quarter and Full Year 2024 Results

work page 2025
[2]

American Airlines Group Inc. (2025). Fourth-Quarter and Full-Year 2024 Results. Form 8-K, Exhibit 99.1

work page 2025
[3]

S., & Ghosh, J

Banerjee, A., Merugu, S., Dhillon, I. S., & Ghosh, J. (2005). Clustering with Bregman Diver- gences. Journal of Machine Learning Research, 6, 1705–1749

work page 2005
[4]

Beverage Digest. (2025). Fact Book, 30th Edition

work page 2025
[5]

Biderman, S., et al. (2023). Pythia: A Suite for Analyzing Large Language Models Across Training and Scaling. arXiv:2304.01373

work page internal anchor Pith review Pith/arXiv arXiv 2023
[6]

(2024, December 17)

Bloomberg. (2024, December 17). Monster, Red Bull on the Run With Energy Drinks Going Sugar-Free

work page 2024
[7]

Box Office Mojo. (2025). Domestic Box Office for 2024

work page 2025
[8]

A., & Terry, M

Bradley, R. A., & Terry, M. E. (1952). Rank Analysis of Incomplete Block Designs: I. The Method of Paired Comparisons. Biometrika, 39(3/4), 324–345. 22

work page 1952
[9]

Bregman, L. M. (1967). The Relaxation Method of Finding the Common Point of Convex Sets and Its Application to the Solution of Problems in Convex Programming. USSR Computational Mathematics and Mathematical Physics, 7(3), 200–217

work page 1967
[10]

Celsius Holdings Inc. (2025). Celsius Holdings Reports Fourth Quarter and Full Year 2024 Financial Results

work page 2025
[11]

Cerebras Systems. (2023). cerebras/Cerebras-GPT-590M. Hugging Face

work page 2023
[12]

Chen, R., et al. (2025). Persona Vectors: Monitoring and Controlling Character Traits in Lan- guage Models. arXiv:2507.21509

work page internal anchor Pith review Pith/arXiv arXiv 2025
[13]

Delta Air Lines. (2025). Delta Air Lines Announces December Quarter and Full Year 2024 Financial Results

work page 2025
[14]

Dey, N., Gosal, G., Chen, Z., Khachane, H., Marshall, W., Pathria, R., Tom, M., & Hestness, J. (2023). Cerebras-GPT: Open Compute-Optimal Language Models Trained on the Cerebras Wafer-Scale Cluster. arXiv:2304.03208

work page arXiv 2023
[15]

EleutherAI. (2023). EleutherAI/pythia-160m. Hugging Face

work page 2023
[16]

M., & Schindelin, J

Endres, D. M., & Schindelin, J. E. (2003). A New Metric for Probability Distributions. IEEE Transactions on Information Theory, 49(7), 1858–1860

work page 2003
[17]

Evidnt. (2025). Energy Drink Market Trends 2025: Brand Leaders, Retail Insights, Functional Innovation

work page 2025
[18]

Federal Aviation Administration. (2025). Calendar Year 2024 Commercial Service Enplane- ments

work page 2025
[19]

Fein, A. J. (2025, March 11). The Top 15 U.S. Pharmacies of 2024: Market Shares and Rev- enues at the Biggest Chains, PBMs, and Specialty Pharmacies. Drug Channels

work page 2025
[20]

Gao, L., et al. (2021). The Pile: An 800GB Dataset of Diverse Text for Language Modeling. arXiv:2101.00027

work page internal anchor Pith review Pith/arXiv arXiv 2021
[21]

Gärdenfors, P. (2000). Conceptual Spaces: The Geometry of Thought. MIT Press

work page 2000
[22]

Google. (2025). Gemini 2.0 Flash Model Card

work page 2025
[23]

Google. (2025). Gemini Models. Google AI for Developers

work page 2025
[24]

Google. (2025). Gemma 3 Model Card. Google AI for Developers

work page 2025
[25]

Google. (2025). google/gemma-3-270m. Hugging Face

work page 2025
[26]

Good, P. (2005). Permutation, Parametric and Bootstrap Tests of Hypotheses. Springer. 3rd Edition

work page 2005
[27]

Guo, C., Pleiss, G., Sun, Y ., & Weinberger, K. Q. (2017). On Calibration of Modern Neural Networks. Proceedings of ICML, 1321–1330

work page 2017
[28]

Hernandez, D., et al. (2022). Scaling Laws and Interpretability of Learning from Repeated Data. arXiv:2205.10487

work page internal anchor Pith review Pith/arXiv arXiv 2022
[29]

W., Smale, S., & Devaney, R

Hirsch, M. W., Smale, S., & Devaney, R. L. (2012). Differential Equations, Dynamical Sys- tems, and an Introduction to Chaos. Academic Press. 3rd Edition

work page 2012
[30]

Holtzman, A., Buys, J., Du, L., Forbes, M., & Choi, Y . (2020). The Curious Case of Neural Text Degeneration. Proceedings of ICLR

work page 2020
[31]

Hunter, D. R. (2004). MM Algorithms for Generalized Bradley–Terry Models. Annals of Statistics, 32(1), 384–406

work page 2004
[32]

Insurance Information Institute. (2025). Facts + Statistics: Insurance company rankings. 23

work page 2025
[33]

Investor’s Business Daily. (2025). Celsius Stock Pops on Energy Drink Market Momentum

work page 2025
[34]

JetBlue Airways Corporation. (2025). Annual Report on Form 10-K for fiscal year ended De- cember 31, 2024

work page 2025
[35]

Kelly, G. M. (1982). Basic Concepts of Enriched Category Theory. Cambridge University Press

work page 1982
[36]

Kullback, S., & Leibler, R. A. (1951). On Information and Sufficiency. Annals of Mathematical Statistics, 22(1), 79–86

work page 1951
[37]

Lawvere, F. W. (1973). Metric Spaces, Generalized Logic, and Closed Categories. Rendiconti del Seminario Matematico e Fisico di Milano, 43, 135–166

work page 1973
[38]

Lin, J. (1991). Divergence Measures Based on the Shannon Entropy. IEEE Transactions on Information Theory, 37(1), 145–151

work page 1991
[39]

Lindsey, J., et al. (2025). On the Biology of a Large Language Model. Transformer Circuits

work page 2025
[40]

Luce, R. D. (1959). Individual Choice Behavior: A Theoretical Analysis. Wiley

work page 1959
[41]

National Retail Federation. (2025). Top 100 Retailers 2025 List

work page 2025
[42]

Ng, K.-S. (2025). On the Definition of Intelligence. arXiv:2507.22423

work page arXiv 2025
[43]

OpenAI. (2024). GPT-4o mini: Advancing Cost-Efficient Intelligence

work page 2024
[44]

OpenAI. (2024). GPT-4o System Card

work page 2024
[45]

Restaurant Business. (2025). Top 500 Chains 2025. Restaurant Business Online

work page 2025
[46]

Rimsky, N., et al. (2024). Steering Llama 2 via Contrastive Activation Addition. Proceedings of ACL, 15387–15405

work page 2024
[47]

Rosch, E. (1975). Cognitive Representations of Semantic Categories. Journal of Experimental Psychology: General, 104(3), 192–233

work page 1975
[48]

Rosch, E., & Mervis, C. B. (1975). Family Resemblances: Studies in the Internal Structure of Categories. Cognitive Psychology, 7(4), 573–605

work page 1975
[49]

Salton, G., Wong, A., & Yang, C.-S. (1975). A Vector Space Model for Automatic Indexing. Communications of the ACM, 18(11), 613–620

work page 1975
[50]

Spearman, C. (1904). The Proof and Measurement of Association between Two Things. The American Journal of Psychology, 15(1), 72–101

work page 1904
[51]

Southwest Airlines Co. (2025). Southwest Airlines Reports Fourth Quarter and Full Year 2024 Results

work page 2025
[52]

Steering Language Models With Activation Engineering

Turner, A. M., et al. (2023). Steering Language Models with Activation Engineering. arXiv:2308.10248

work page internal anchor Pith review Pith/arXiv arXiv 2023
[53]

Tversky, A. (1977). Features of Similarity. Psychological Review, 84(4), 327–352

work page 1977
[54]

United Airlines. (2025). United Airlines 2024 Corporate Responsibility Report

work page 2025
[55]

Zadeh, L. A. (1965). Fuzzy Sets. Information and Control, 8(3), 338–353

work page 1965
[56]

Zou, A., et al. (2023). Representation Engineering: A Top-Down Approach to AI Trans- parency. arXiv:2310.01405. 24

work page internal anchor Pith review Pith/arXiv arXiv 2023