Pith. sign in

REVIEW 3 major objections 5 minor 63 references

A Rose by Any Other Name Would Smell as Sweet: Categorical Homotopy Theory for Large Language Models

T0 review · 3 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read The paper argues that the next-token distributions of a large language model form a Markov category, and that meaning-preserving paraphrases can be treated as weak equivalences in a Quillen model structure on that category.

desk verdict The homotopy framing for paraphrase equivalence is fresh, but Theorem 12's proof rests on a false Kan-complex claim; reject, though the underlying question is worth pursuing. read the letter →

arxiv 2508.10018 v1 pith:6IKAHZVX submitted 2025-08-07 cs.CL cs.AImath.AT

Sridhar Mahadevan This is my paper
classification cs.CLcs.AImath.AT MSC 18N4018M1068T50
keywords categoricalhomotopylargelanguagemodelsMarkovcategoriesmodelKancomplexesparaphraseequivalencesimplicialsetsalgebraicK-theory
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 sets out to give large language models a homotopy theory. Its starting observation is that an LLM should assign the same next-token probabilities to rephrasings such as “Charles Darwin wrote” and “Charles Darwin is the author of,” but in practice it does not. The author proposes to model the LLM as a Markov category—a symmetric monoidal category whose arrows are probability distributions—and to declare meaning-preserving paraphrases to be weak equivalences rather than isomorphisms. The paper's key claim, Theorem 12, is that LLM Markov categories define model categories, so the usual categorical notion of sameness is replaced by a homotopy notion in which paraphrases become isomorphic in the homotopy category. If the claim holds, sentence equivalence becomes a topological invariant, giving a principled basis for comparing and smoothing next-token distributions.

What carries the argument

The load-bearing construction is the nerve of an LLM Markov category, converted into a simplicial set. An LLM Markov category is a symmetric monoidal category with a copy/delete comonoid on each object, so its arrows are probability distributions over token sequences. The nerve sends this category to a simplicial set whose $n$-simplices are composable $n$-tuples of token sequences; the proof that this simplicial set is a Kan complex is the step that makes Theorem 12 work. The paper claims each $(n-1)$-simplex is a retract of an $n$-simplex via a face operator, so all horns fill, and the standard Quillen model structure on simplicial sets transfers to LLM Markov categories.

What would settle it

Fix a real LLM, form its Markov category from next-token probabilities, and search for a horn $\Lambda^3_1$ in the nerve whose three 2-dimensional faces are paraphrases with positive probability but whose missing face has probability zero; such a horn would fail to fill, contradicting the Kan-complex premise and with it Theorem 12.

Watch

Extended reading notes

Core claim

The claim is that the category generated by an LLM's next-token distributions carries a Quillen model structure. Each sentence prefix is an object and each conditional distribution $P(y|x)$ is an arrow in an LLM Markov category; a phrase and its paraphrase, such as “Charles Darwin wrote” and “Charles Darwin is the author of,” are connected not by an isomorphism but by a weak equivalence. Theorem 12 asserts the three classes of arrows: fibrations are Kan complexes, cofibrations are monomorphisms of the simplicial sets obtained by the nerve construction, and weak equivalences are maps whose geometric realizations are homotopy equivalences. The paper develops this through classifying spaces, H-

Load-bearing premise

The argument stands on the assertion that the nerve of an LLM Markov category is a Kan complex—that every horn built from token sequences has a filler—because its n-simplices are just n-token sequences; if a horn can be assembled from probable paraphrases whose composite is improbable or ungrammatical, the model-category proof collapses.

Editorial extensions

If this is right

  • In the homotopy category $\mathrm{Ho}(\mathcal{L})$, paraphrases become isomorphic, so meaning-equivalent sentences can be substituted freely in next-token computations.
  • Semantic equivalence classes are exactly the path components $\pi_0(B\mathcal{L})$ of the classifying space, giving a computable invariant for a language model.
  • The Grayson–Quillen construction $\mathcal{L}^{-1}\mathcal{L}$ assembles equivalence classes into an abelian group $K_0(\mathcal{L})$, enabling algebraic operations on meanings.
  • If LLM Markov categories are model categories, the lifting-problem machinery applies to language: comparing two LLMs means studying maps between their model categories that preserve cofibrations, fibrations, and weak equivalences.

Reading between the lines

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

  • The Kan-complex premise is the most empirically checkable part: it predicts that every horn composed of probable token sequences has a probable filler, so a search for unfillable horns on a real LLM's next-token distributions would directly test Theorem 12.
  • The framework implies a stronger invariance principle than k-NN smoothing: a paraphrase-invariant language model would be a functor out of the localized category $\mathcal{L}(\Sigma^{-1})$, giving a universal property that any smoothing method ought to satisfy.
  • Because weak equivalences are defined through geometric realization, the framework predicts that semantic similarity is continuous under small perturbations of token sequences, a claim that can be probed by measuring how homotopy classes change along token-space paths.
  • If the model structure exists only for abstract Markov categories and not for the finite, approximate distributions of deployed LLMs, the practical payoff would vanish; the retract argument in the proof sketch is where that gap would show up.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes an abstract categorical-homotopy framework for LLM paraphrase equivalence. It introduces 'LLM Markov categories' (Definition 18), reviews enriched categories, ends/coends, Markov categories, simplicial sets, Kan complexes, and model categories, and claims as its central result that LLM Markov categories define model categories (Theorem 12, Section 7.4). The intended application is to identify semantically equivalent sentences through weak equivalences, motivated by next-token distribution discrepancies for paraphrases such as 'Charles Darwin wrote' versus 'Charles Darwin is the author of'.

Significance. The motivating problem is real, and drawing on Markov categories and simplicial methods is a plausible source of useful abstractions. The expository portions collect standard material in one place, and the paper correctly identifies that paraphrase equivalence is a homotopy-theoretic phenomenon. However, the central theorem is not established: the proof rests on a false criterion for Kan complexes and never defines the model structure. No computational or empirical application is offered. If Theorem 12 were valid it would be a notable contribution to categorical NLP; in its present form the paper is best read as a research proposal.

major comments (3)
  1. [§7.4, Theorem 12] The proof asserts that the nerve of an LLM Markov category is a Kan complex because each n-simplex is a sequence of n tokens and each (n−1)-simplex is a retract of an n-simplex by a face operator. Neither observation is the Kan condition, which requires fillers for every horn Λ^n_i → X. For any small category C, N(C) is Kan iff C is a groupoid. An LLM Markov category (Definition 18) is not generally a groupoid: next-token extension morphisms are not invertible. Concretely, for a non-invertible f:A→B, the horn Λ^2_0 consisting of f:A→B and id_A:A→A has no filler unless there is a morphism B→A. Section 7 itself concedes that ordinary categorical nerves are only quasi-categories; no argument supplies the needed groupoid structure. The proof also misstates the role of Kan complexes, calling them fibrations rather than fibrant objects.
  2. [§7.4, Theorem 12; §7.3, MC1–MC5] Even if N(C) were Kan, that fact only makes N(C) a fibrant object in the model category of simplicial sets; it does not produce a model structure on the original category C. The theorem states that the LLM Markov category itself carries a model structure, but the weak equivalences are never defined: Section 9.1's 'topological embeddings' need not be weak homotopy equivalences. MC1 (all small limits and colimits) is not verified for the LLM Markov category, and MC2–MC5 are not checked. The appeal to 'simplicial sets define model categories' does not transfer without an explicit adjunction or Quillen equivalence.
  3. [Definition 27] The paper defines a homotopy between morphisms f0, f1 : X → Y in an LLM Markov category as a morphism h : Δ^1 × X → Y. Markov categories are not necessarily Cartesian, so the product Δ^1 × X need not exist, and Δ^1 is not specified as an object of the category. The homotopy notion on which the framework depends is therefore not well-defined as stated.
minor comments (5)
  1. [Abstract and §1] The motivating paraphrase example switches between 'Charles Darwin wrote' and 'Charles Dickens wrote' in the abstract and Section 1. Use one consistent example.
  2. [Definition 15] The associator has a typo: the target is written as (C1⊗C2)⊗C2 but should be (C1⊗C2)⊗C3.
  3. [Definition 18] The definition says every object is a token, while Definitions 4–5 and Section 3.3 treat sentence fragments as objects. Clarify the relationship between tokens, fragments, and hom-objects that carry next-token probabilities.
  4. [Theorem 16] The category CGHaus is normally defined as compactly generated weak Hausdorff spaces, not merely compact Hausdorff spaces with compactly continuous mappings. As stated, the theorem is misleading.
  5. [Theorem 12] Typo 'emebddings' should be 'embeddings.' Also, the proof sketch's three bullets do not correspond to any of the model-category axioms MC1–MC5.

Circularity Check

2 steps flagged · score 6.0 of 10

Theorem 12 reduces the model-category claim to an unproved Kan-complex assertion; a side theorem restates its own definition.

  1. self definitional [Section 7.4, Theorem 12 and proof sketch]
    "Simplicial sets constructed using the nerve functor over LLM Markov categories are Kan complexes as each n-simplex is a sequence of n-length tokens. Any n−1 simplex Xn−1 of an LLM is a retract of an n-simplex Xn of an LLM by applying a face operator (see Examples 2-4)."

    Theorem 12's conclusion is that LLM Markov categories define model categories. The proof's only substantive step is to assert that the nerve of such a category is a Kan complex. The paper's own Section 7 concedes that ordinary categorical nerves are only quasi-categories and 'do not in general lead to model categories'; the Kan condition is exactly the additional structure that would make the nerve a fibrant model-category object. Instead of deriving horn-filling, the proof re-describes the nerve ('each n-simplex is a sequence of n-length tokens') and adds that an (n−1)-simplex is a retract of an n-simplex under a face operator—neither fact implies the Kan property. The weak equivalences are never defined independently (they are only 'associated with topological embeddings'), so the claime

  2. self definitional [Section 6.2, Definition 35 and Theorem 11]
    "Definition 35. Define the LLM groupoid as the category LG whose objects are defined as the equivalence classes of the connected paths of the category, and whose invertible morphisms correspond to invertible edges that map from an equivalence class back to itself. Theorem 11. The classifying space of the LLM groupoid category BLG is defined as BLG = ⨆_i BL_i^G where disjoint sum index i ranges over equivalence classes."

    Theorem 11 is labelled a theorem but its assertion is exactly the content of Definition 35: LG is defined by equivalence classes of connected paths, and then its classifying space is 'proved' to decompose as a disjoint union over equivalence classes. No independent construction or universal property is provided; the 'proof' only cites a textbook result on classifying spaces of groupoids. This is a definitional tautology. It is not load-bearing for the main model-category claim, but it is a genuine instance of a claimed result that reduces to its own definition.

full rationale

The paper's surrounding results (Theorems 7–11, K-theory constructions, classifying-space statements) are mostly citations to standard external textbooks (Richter, Segal, Quillen, Hovey) and are not circular in themselves. The load-bearing problem is Theorem 12: the proof collapses the entire model-category claim into the assertion that the nerve of an LLM Markov category is Kan, without proving horn-filling and without specifying or verifying the weak equivalences required by Definition 42 (MC1–MC5). The paper itself flags 'number of tricky issues here to resolve, which lie beyond the scope of this introductory paper' immediately before the theorem, further indicating that the derivation is incomplete rather than self-contained. This is a serious correctness gap, but it is also a partial circularity: the model-category conclusion is being assumed under the guise of the Kan-complex premise. The Definition 35/Theorem 11 pair is a minor additional definitional tautology. There is no load-bearing self-citation; the only self-reference (Mahadevan 2023) appears in a list of categorical-causality references and is not used to justify the central claim. Because the central claim partially reduces by construction to an assumed property, the score is 6.

Assumptions & free parameters 0 free parameters · 7 assumptions · 3 invented entities

No numerical fits are performed, so there are no fitted constants. The load-bearing choices are structural: that LLM outputs form a Markov category, that the nerve is Kan, that MC1 holds, and that paraphrases can be modeled as weak equivalences. These are recorded as axioms or ad hoc assumptions because none is derived from machine-checked proofs, shipped code, or empirical tests.

assumptions (7)
  • standard math The Quillen model structure on simplicial sets exists and can be imported.
    Theorem 12's proof says it is a special case of the result that simplicial sets define model categories, relying on Quillen, Hovey, and May-Ponto.
  • standard math The nerve functor turns a category into a simplicial set and classifies its path components.
    Theorems 7 through 11 rely on the standard nerve/classifying space results of Segal and Richter.
  • domain assumption LLM next-token probability distributions form the morphisms of a Markov category.
    Definition 18 stipulates objects are tokens and arrows are probabilities; this modeling choice is not derived from transformer architecture or trained LLM behavior.
  • ad hoc to paper The nerve of an LLM Markov category is a Kan complex.
    Section 7.4 proof of Theorem 12 asserts this, but the stated retraction argument does not imply horn filling, and ordinary categorical nerves are not Kan unless the category is a groupoid.
  • ad hoc to paper An LLM Markov category has all small limits and colimits, satisfying MC1.
    Definition 42 requires MC1 for a model category, but no construction or proof of limits and colimits is given for C_LLM.
  • ad hoc to paper Weak equivalences of an LLM model category are meaningfully associated with topological embeddings of simplicial sets.
    Section 7.4 states this association without defining which maps count as weak equivalences, so the claimed model structure cannot be verified.
  • domain assumption Paraphrases such as "Charles Darwin wrote" and "Charles Darwin is the author of" should be treated as the same or equivalent for next-token prediction.
    The Introduction and Figure 1 take this as the motivating desideratum. It is asserted as a property of meaning rather than tested empirically.
invented entities (3)
  • LLM Markov category C_LLM
    purpose: To encode LLM next-token probabilities as arrows in a symmetric monoidal category with copy and delete structure.
    Definition 18 is a relabeling of a Markov category with token objects; no falsifiable handle connects it to actual transformer weights or next-token probabilities.
  • LLM model category structure
    purpose: To make paraphrase equivalence a formal weak equivalence and enable homotopy-theoretic comparison of sentence fragments.
    Theorem 12 claims existence but does not concretely define weak equivalences, fibrations, and cofibrations, nor verify MC1 through MC5.
  • LLM groupoid L_G
    purpose: To represent equivalence classes of paraphrases as a groupoid whose classifying space splits over equivalence classes.
    Definition 35 depends on equivalence classes that are never independently characterized, so the object has no external predictive content.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Rose by Any Other Name Would Smell as Sweet: Categorical Homotopy Theory for Large Language Models." pith.science (2026). https://pith.science/paper/6IKAHZVX

@misc{pith2026250810018,
  author       = {Pith},
  title        = {Pith review of: A Rose by Any Other Name Would Smell as Sweet: Categorical Homotopy Theory for Large Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6IKAHZVX}},
  note         = {Machine review of arXiv:2508.10018}
}
read the original abstract

Natural language is replete with superficially different statements, such as ``Charles Darwin wrote" and ``Charles Darwin is the author of", which carry the same meaning. Large language models (LLMs) should generate the same next-token probabilities in such cases, but usually do not. Empirical workarounds have been explored, such as using k-NN estimates of sentence similarity to produce smoothed estimates. In this paper, we tackle this problem more abstractly, introducing a categorical homotopy framework for LLMs. We introduce an LLM Markov category to represent probability distributions in language generated by an LLM, where the probability of a sentence, such as ``Charles Darwin wrote" is defined by an arrow in a Markov category. However, this approach runs into difficulties as language is full of equivalent rephrases, and each generates a non-isomorphic arrow in the LLM Markov category. To address this fundamental problem, we use categorical homotopy techniques to capture ``weak equivalences" in an LLM Markov category. We present a detailed overview of application of categorical homotopy to LLMs, from higher algebraic K-theory to model categories, building on powerful theoretical results developed over the past half a century.

Figures

Figures reproduced from arXiv: 2508.10018 by the authors.

Figure 1
Figure 1. A lifting diagram for LLMs: any such commutative diagram has a “lift" [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Structure of monoidal categories. 4 Monoidal LLM Categories In this section, we present a categorical framework for analyzing LLMs based on a symmetric monoidal category representation called a Markov category [Fritz, 2020]. We first give a brief explanation of symmetric monoidal categories, before more formally introducing Markov categories. We compare our approach to previous work on categorical representations of… view at source ↗
Figure 3
Figure 3. Homotopy theory models equivalence classes in an LLM category. [PITH_FULL_IMAGE:figures/full_fig_p015_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: A simplicial set [May, 1992] representation of an LLM consists of a graded sequence of sets, where [PITH_FULL_IMAGE:figures/full_fig_p017_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

63 extracted references · 38 canonical work pages

  1. [2]

    Codensity and the giry monad

    Tom Avery. Codensity and the giry monad. Journal of Pure and Applied Algebra, 220 0 (3): 0 1229–1251, March 2016. ISSN 0022-4049. doi:10.1016/j.jpaa.2015.08.017. URL http://dx.doi.org/10.1016/j.jpaa.2015.08.017

    work page doi:10.1016/j.jpaa.2015.08.017 2016

  2. [3]

    Boardman and Rainer Vogt

    M. Boardman and Rainer Vogt. Homotopy invariant algebraic structures on topological spaces. Springer, Berlin, 1973

    work page 1973

  3. [4]

    Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S

    Rishi Bommasani, Drew A. Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S. Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, Erik Brynjolfsson, Shyamal Buch, Dallas Card, Rodrigo Castellon, Niladri Chatterji, Annie Chen, Kathleen Creel, Jared Quincy Davis, Dora Demszky, Chris Donahue, Moussa Doumbouya, Esin Durmus, Stef...

    work page 2022

  4. [5]

    Handbook of Categorical Algebra, volume 2 of Encyclopedia of Mathematics and its Applications

    Francis Borceux. Handbook of Categorical Algebra, volume 2 of Encyclopedia of Mathematics and its Applications. Cambridge University Press, 1994. doi:10.1017/CBO9780511525858

    doi:10.1017/cbo9780511525858 1994

  5. [6]

    Bradley, J

    TD. Bradley, J. Terilla, and Y. Vlassopoulos. An enriched category theory of language: From syntax to semantics. La Matematica, 1: 0 551--580, 2022

    work page 2022

  6. [7]

    An attentive survey of attention models, 2021

    Sneha Chaudhari, Varun Mithal, Gungor Polatkan, and Rohan Ramanath. An attentive survey of attention models, 2021. URL https://arxiv.org/abs/1904.02874

    arXiv 2021

  7. [8]

    Disintegration and bayesian inversion via string diagrams

    Kenta Cho and Bart Jacobs. Disintegration and bayesian inversion via string diagrams. Mathematical Structures in Computer Science, 29 0 (7): 0 938–971, March 2019. ISSN 1469-8072. doi:10.1017/s0960129518000488. URL http://dx.doi.org/10.1017/S0960129518000488

    doi:10.1017/s0960129518000488 2019

  8. [9]

    The Mathematics of Text Structure

    Bob Coecke. The mathematics of text structure. ArXiv, abs/1904.03478, 2019. URL https://api.semanticscholar.org/CorpusID:102351346

    work page Pith review arXiv 1904

Show all 63 references
  1. [10]

    The mathematics of text structure, 2020

    Bob Coecke. The mathematics of text structure, 2020

    2020

  2. [11]

    Picturing Quantum Processes: A First Course in Quantum Theory and Diagrammatic Reasoning

    Bob Coecke and Aleks Kissinger. Picturing Quantum Processes: A First Course in Quantum Theory and Diagrammatic Reasoning. Cambridge University Press, 2017

    2017

  3. [12]

    Mathematical foundations for a compositional distributional model of meaning, 2010

    Bob Coecke, Mehrnoosh Sadrzadeh, and Stephen Clark. Mathematical foundations for a compositional distributional model of meaning, 2010. URL https://arxiv.org/abs/1003.4394

    2010 arXiv

  4. [13]

    DeepSeek-AI, Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, Xiaokang Zhang, Xingkai Yu, Yu Wu, Z. F. Wu, Zhibin Gou, Zhihong Shao, Zhuoshu Li, Ziyi Gao, Aixin Liu, Bing Xue, Bingxuan Wang, Bochao Wu, Bei F...

    2025 arXiv

  5. [14]

    Logits of API-Protected LLMs leak proprietary information, 2024

    Matthew Finlayson, Xiang Ren, and Swabha Swayamdipta. Logits of API-Protected LLMs leak proprietary information, 2024. URL https://arxiv.org/abs/2403.09539

    2024 arXiv

  6. [15]

    Causal theories: A categorical perspective on bayesian networks

    Brendan Fong. Causal theories: A categorical perspective on bayesian networks. Master's thesis, Oxford University, 2012

    2012

  7. [16]

    Seven Sketches in Compositionality: An Invitation to Applied Category Theory

    Brendan Fong and David I Spivak. Seven Sketches in Compositionality: An Invitation to Applied Category Theory. Cambridge University Press, 2018

    2018

  8. [17]

    Spivak, and R \' e my Tuy \' e ras

    Brendan Fong, David I. Spivak, and R \' e my Tuy \' e ras. Backprop as functor: A compositional perspective on supervised learning. In 34th Annual ACM/IEEE Symposium on Logic in Computer Science, LICS 2019, Vancouver, BC, Canada, June 24-27, 2019 , pages 1--13. IEEE , 2019. do...

    2019

  9. [18]

    Universal Coalgebras

    James Fox. Universal Coalgebras. Dissertation, McGill University, 1976

    1976

  10. [19]

    A synthetic approach to markov kernels, conditional independence and theorems on sufficient statistics

    Tobias Fritz. A synthetic approach to markov kernels, conditional independence and theorems on sufficient statistics. Advances in Mathematics, 370: 0 107239, August 2020. ISSN 0001-8708. doi:10.1016/j.aim.2020.107239. URL http://dx.doi.org/10.1016/j.aim.2020.107239

    2020

  11. [20]

    The d-separation criterion in categorical probability

    Tobias Fritz and Andreas Klingler. The d-separation criterion in categorical probability. Journal of Machine Learning Research, 24 0 (46): 0 1--49, 2023. URL http://jmlr.org/papers/v24/22-0916.html

    2023

  12. [21]

    Gabriel, P

    P. Gabriel, P. Gabriel, and M. Zisman. Calculus of Fractions and Homotopy Theory. Calculus of Fractions and Homotopy Theory. Springer-Verlag, 1967. ISBN 9780387037776. URL https://books.google.com/books?id=UEQZAQAAIAAJ

    1967

  13. [22]

    The unreasonable power of the lifting property in elementary mathematics, 2017

    Misha Gavrilovich. The unreasonable power of the lifting property in elementary mathematics, 2017. URL https://arxiv.org/abs/1707.06615

    2017 arXiv

  14. [23]

    Higher algebraic k-theory: Ii

    Daniel Grayson. Higher algebraic k-theory: Ii. In Michael R. Stein, editor, Algebraic K-Theory, pages 217--240, Berlin, Heidelberg, 1976. Springer Berlin Heidelberg. ISBN 978-3-540-37964-5

    1976

  15. [24]

    Efficiently modeling long sequences with structured state spaces

    Albert Gu, Karan Goel, and Christopher R \' e . Efficiently modeling long sequences with structured state spaces. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022 . OpenReview.net, 2022. URL https://openreview.net/f...

    2022

  16. [25]

    Theoretical limitations of self-attention in neural sequence models

    Michael Hahn. Theoretical limitations of self-attention in neural sequence models. Trans. Assoc. Comput. Linguistics, 8: 0 156--171, 2020. doi:10.1162/TACL\_A\_00306. URL https://doi.org/10.1162/tacl\_a\_00306

    2020 doi

  17. [26]

    Measure Theory

    Paul Halmos. Measure Theory. Springer, 1974

    1974

  18. [27]

    Categories for Quantum Theory: An Introduction

    Chris Heunen and Jamie Vicary. Categories for Quantum Theory: An Introduction. Oxford University Press, November 2019. doi:10.1093/oso/9780198739623.001.0001

    2019

  19. [28]

    Model categories

    Mark Hovey. Model categories. Foundations of Stable Homotopy Theory, 2020. URL https://api.semanticscholar.org/CorpusID:241323754

    2020

  20. [29]

    Distilling step-by-step! outperforming larger language models with less training data and smaller model sizes, 2023

    Cheng-Yu Hsieh, Chun-Liang Li, Chih-Kuan Yeh, Hootan Nakhost, Yasuhisa Fujii, Alexander Ratner, Ranjay Krishna, Chen-Yu Lee, and Tomas Pfister. Distilling step-by-step! outperforming larger language models with less training data and smaller model sizes, 2023. URL https://arxi...

    2023 arXiv

  21. [30]

    Causal inference by string diagram surgery, 2018

    Bart Jacobs, Aleks Kissinger, and Fabio Zanasi. Causal inference by string diagram surgery, 2018. URL https://arxiv.org/abs/1811.08338

    2018 arXiv

  22. [31]

    A. Joyal. Quasi-categories and kan complexes. Journal of Pure and Applied Algebra, 175 0 (1): 0 207--222, 2002. ISSN 0022-4049. doi:https://doi.org/10.1016/S0022-4049(02)00135-4. URL https://www.sciencedirect.com/science/article/pii/S0022404902001354. Special Volume celebratin...

    2002 doi

  23. [32]

    lambeq: An efficient high-level python library for quantum nlp, 2021

    Dimitri Kartsaklis, Ian Fan, Richie Yeung, Anna Pearson, Robin Lorenz, Alexis Toumi, Giovanni de Felice, Konstantinos Meichanetzidis, Stephen Clark, and Bob Coecke. lambeq: An efficient high-level python library for quantum nlp, 2021. URL https://arxiv.org/abs/2110.04236

    2021 arXiv

  24. [33]

    Generalization through memorization: Nearest neighbor language models, 2020

    Urvashi Khandelwal, Omer Levy, Dan Jurafsky, Luke Zettlemoyer, and Mike Lewis. Generalization through memorization: Nearest neighbor language models, 2020. URL https://arxiv.org/abs/1911.00172

    2020 arXiv

  25. [34]

    Dan Klein and Christopher D. Manning. A * parsing: Fast exact V iterbi parse selection. In Proceedings of the 2003 Human Language Technology Conference of the North A merican Chapter of the Association for Computational Linguistics , pages 119--126, 2003. URL https://aclanthol...

    2003

  26. [35]

    J. Lambek. Pregroups and natural language processing. The Mathematical Intelligencer, 28: 0 41--48, 2006

    2006

  27. [36]

    A * CCG parsing with a supertag-factored model

    Mike Lewis and Mark Steedman. A * CCG parsing with a supertag-factored model. In Alessandro Moschitti, Bo Pang, and Walter Daelemans, editors, Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing ( EMNLP ) , pages 990--1000, Doha, Qatar, Octob...

    2014 doi

  28. [37]

    (Co)end Calculus

    Fosco Loregian. (Co)end Calculus. London Mathematical Society Lecture Note Series. Cambridge University Press, 2021. doi:10.1017/9781108778657

    2021 doi

  29. [38]

    Higher Topos Theory

    Jacob Lurie. Higher Topos Theory . Annals of mathematics studies. Princeton University Press, Princeton, NJ, 2009. URL https://cds.cern.ch/record/1315170

    2009

  30. [39]

    Sheaves in Geometry and Logic a First Introduction to Topos Theory

    Saunders Mac Lane and Ieke Moerdijk. Sheaves in Geometry and Logic a First Introduction to Topos Theory. Springer New York, New York, NY, 1992. ISBN 9781461209270 1461209277. URL http://link.springer.com/book/10.1007/978-1-4612-0927-0

    1992 doi

  31. [40]

    Categories for the Working Mathematician

    Saunders MacLane. Categories for the Working Mathematician. Springer-Verlag, New York, 1971. Graduate Texts in Mathematics, Vol. 5

    1971

  32. [41]

    Universal causality

    Sridhar Mahadevan. Universal causality. Entropy, 25 0 (4): 0 574, 2023. doi:10.3390/E25040574. URL https://doi.org/10.3390/e25040574

    2023 doi

  33. [42]

    J.P. May. Simplicial Objects in Algebraic Topology. University of Chicago Press, 1992

    1992

  34. [43]

    May and K

    J.P. May and K. Ponto. More Concise Algebraic Topology: Localization, Completion, and Model Categories. Chicago Lectures in Mathematics. University of Chicago Press, 2012. ISBN 9780226511788. URL https://books.google.com/books?id=SHhmxUPskFwC

    2012

  35. [44]

    Umap: Uniform manifold approximation and projection for dimension reduction, 2018

    Leland McInnes, John Healy, and James Melville. Umap: Uniform manifold approximation and projection for dimension reduction, 2018. URL https://arxiv.org/abs/1802.03426

    2018 arXiv

  36. [45]

    William Merrill, Ashish Sabharwal, and Noah A. Smith. Saturated transformers are constant-depth threshold circuits, 2022. URL https://arxiv.org/abs/2106.16213

    2022 arXiv

  37. [46]

    Can you learn semantics through next-word prediction? the case of entailment, 2024

    William Merrill, Zhaofeng Wu, Norihito Naka, Yoon Kim, and Tal Linzen. Can you learn semantics through next-word prediction? the case of entailment, 2024. URL https://arxiv.org/abs/2402.13956

    2024 arXiv

  38. [47]

    The geometric realization of a semi-simplicial complex

    John Milnor. The geometric realization of a semi-simplicial complex. The Annals of Mathematics, 65 0 (2): 0 357--362, 1957

    1957

  39. [48]

    Levels of AGI

    Meredith Ringel Morris, Jascha Sohl-Dickstein, Noah Fiedel, Tris Warkentin, Allan Dafoe, Aleksandra Faust, Clement Farabet, and Shane Legg, editors. Levels of AGI for Operationalizing Progress on the Path to AGI, 2023. Original arXiv title in November 2023 was "Levels of AGI":...

    2023

  40. [49]

    James R. Munkres. Elements of algebraic topology. Addison-Wesley, 1984. ISBN 978-0-201-04586-4

    1984

  41. [50]

    Higher algebraic k-theory: I

    Daniel Quillen. Higher algebraic k-theory: I. In H. Bass, editor, Higher K-Theories, pages 85--147, Berlin, Heidelberg, 1973. Springer Berlin Heidelberg. ISBN 978-3-540-37767-2

    1973

  42. [51]

    Daniel G. Quillen. Homotopical algebra. Springer, 1967

    1967

  43. [52]

    B. Richter. From Categories to Homotopy Theory. Cambridge Studies in Advanced Mathematics. Cambridge University Press, 2020. ISBN 9781108479622. URL https://books.google.com/books?id=pnzUDwAAQBAJ

    2020

  44. [53]

    Infinity category theory from scratch, 2019

    Emily Riehl and Dominic Verity. Infinity category theory from scratch, 2019. URL https://arxiv.org/abs/1608.05314

    2019 arXiv

  45. [54]

    Classifying spaces and spectral sequences

    Graeme Segal. Classifying spaces and spectral sequences. Mathématiques de l’Institut des Hautes Scientifiques, 34: 0 92--100, 1968

    1968

  46. [55]

    Selinger

    P. Selinger. A survey of graphical languages for monoidal categories. In New Structures for Physics, pages 289--355. Springer Berlin Heidelberg, 2010. doi:10.1007/978-3-642-12821-9_4. URL https://doi.org/10.1007

    2010 doi

  47. [56]

    Self-attention with relative position representations, 2018

    Peter Shaw, Jakob Uszkoreit, and Ashish Vaswani. Self-attention with relative position representations, 2018. URL https://arxiv.org/abs/1803.02155

    2018 arXiv

  48. [57]

    Generative modeling by estimating gradients of the data distribution

    Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. In Hanna M. Wallach, Hugo Larochelle, Alina Beygelzimer, Florence d'Alch \' e - Buc, Emily B. Fox, and Roman Garnett, editors, Advances in Neural Information Processing Systems 3...

    2019

  49. [58]

    D. I. Spivak. Metric realization of fuzzy simplicial sets, 2009

    2009

  50. [59]

    David I. Spivak. Database queries and constraints via lifting problems. Mathematical Structures in Computer Science, 24 0 (6), oct 2013. doi:10.1017/s0960129513000479. URL https://doi.org/10.1017

    2013 doi

  51. [60]

    Sutton and Andrew G

    Richard S. Sutton and Andrew G. Barto. Reinforcement learning - an introduction. Adaptive computation and machine learning. MIT Press, 1998. ISBN 978-0-262-19398-6. URL https://www.worldcat.org/oclc/37293240

    1998

  52. [61]

    Gomez, Lukasz Kaiser, and Illia Polosukhin

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Isabelle Guyon, Ulrike von Luxburg, Samy Bengio, Hanna M. Wallach, Rob Fergus, S. V. N. Vishwanathan, and Roman Garnett, ...

    2017

  53. [62]

    N. Yoneda. On ext and exact sequences. J. Fac. Sci. Univ. Tokyo, Sect. I 8: 0 507--576, 1960

    1960

  54. [63]

    A * CCG parsing with a supertag and dependency factored model

    Masashi Yoshikawa, Hiroshi Noji, and Yuji Matsumoto. A * CCG parsing with a supertag and dependency factored model. In Regina Barzilay and Min-Yen Kan, editors, Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pag...

    2017 doi

  55. [64]

    Reddi, and Sanjiv Kumar

    Chulhee Yun, Srinadh Bhojanapalli, Ankit Singh Rawat, Sashank J. Reddi, and Sanjiv Kumar. Are transformers universal approximators of sequence-to-sequence functions? In 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 202...

    2020

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.