Pith. sign in

REVIEW 4 major objections 5 minor 29 references

Why Neural Network Can Discover Symbolic Structures with Gradient-based Training: An Algebraic and Geometric Foundation for Neurosymbolic Reasoning

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Under one geometric symmetry, gradient training solves symbolic constraints one by one

desk verdict The measure-space algebra is a clean extension of Tian, but the main decoupling theorem is internally inconsistent: odd-degree monomials force an even velocity field, which an O(d)-equivariance condition with -I forces to be zero, so the theorem is vacuous as stated. read the letter →

arxiv 2506.21797 v2 pith:7PVWIKSY submitted 2025-06-26 cs.LG

Peihao Wang , Zhangyang Wang This is my paper · ORCID
classification cs.LG
keywords neurosymbolicreasoningWassersteingradientflowmonomialpotentialsgroupequivariancemeasure-valueddynamicsalgebraicstructuresamplecomplexitysymbolicemergence
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

The paper argues that discrete symbolic structure can emerge from continuous gradient-based training as a direct consequence of geometry, not as an accident. Lifting network parameters to a probability measure and treating training as Wasserstein gradient flow, the authors show that when the velocity field is invariant under rotations of parameter space, the algebraic quantities called monomial potentials each evolve independently, each descending toward a binary 0/1 value. The task thus becomes a boolean satisfaction problem in the space of these potentials, and the solution space carries a commutative semi-ring structure whose operations let global optimizers be composed from simpler ones. The same dynamics shrink the effective dimension of the measure to the number of potentials. If the claims hold, this gives a measure-theoretic explanation for how networks internalize rules such as Abelian group addition.

What carries the argument

The load-bearing object is the monomial potential, defined as the expectation of a monomial $r(z)=\prod_{i\in I} z_i$ under the parameter measure $\mu$; the loss is assumed to depend on the measure only through finitely many such potentials. The argument is carried by three mechanisms: a measure-space semi-ring in which addition fuses measures and multiplication pushes forward the elementwise product of samples, with potentials acting as ring homomorphisms; the $\mathrm{O}(d)$-equivariance of the velocity field, which maintains a symmetric measure so that cross-terms $\int \nabla r_i \cdot \nabla r_j\,d\mu_t$ vanish for all time; and a second-variation analysis showing the Hessian's eigenfunctions live in the span of the monomials, making the effective dimension no larger than the number of potentials. Together they turn the infinite-dimensional Wasserstein gradient flow into coordinate-wise descent over a finite set of algebraic variables.

What would settle it

Simulate the two-layer quadratic-activation network on modular addition with complex Fourier parameters and measure the off-diagonal Gram entries $G_{ij}(t)=\int \nabla r_i(z)^\top \nabla r_j(z)\,d\mu_t(z)$. Theorem 4.2 predicts these remain exactly zero throughout training; observing any nonzero value, or constructing a complex-valued parameter vector that makes them nonzero, would falsify the decoupling claim on the paper's own example.

Watch

Extended reading notes

Core claim

The central claim is that, under Gaussian initialization, odd-degree monomial potentials of degree at least three, and an $\mathrm{O}(d)$-equivariant velocity field, the Wasserstein gradient flow decouples coordinate-wise in the space of monomial potentials: $\partial_t \rho_{r_i}[\mu_t] = -C_i(t)\,\partial_{\rho_{r_i}} L(\rho)$, with $C_i(t)>0$. Each potential therefore follows its own scalar gradient, and training drives the vector $\rho$ toward the binary (0/1) assignments that satisfy the logical expression encoded by the loss. The paper further shows that the space of measures with addition (fusing mass) and multiplication (coupling through elementwise product) is a commutative semi-ring, and each monomial potential is a ring homomorphism, so the 0/1-solutions compose algebraically. A companion result states that eigenfunctions of the Hessian operator lie in the span of the monomial set and that eigenvalues are non-increasing under the flow, bounding the effective degrees of freedom by the number of monomial potentials. The upshot is a proof that geometric invariance converts a high-dimensional continuous optimization into a low-dimensional, algebraically structured search over boolean variables.

Load-bearing premise

The proof rests on two unverified premises: all parameters are real numbers even though the motivating group task is naturally complex, and the training dynamics are spherically symmetric even though this is not shown for the task's architecture.

Editorial extensions

If this is right

  • Under an $\mathrm{O}(d)$-invariant loss, each monomial potential reaches its target 0/1 value with exponential convergence, so the network's final state satisfies the task's algebraic constraints rather than merely fitting the training set.
  • Global minimizers compose: multiplying a minimizer by any measure whose potentials are all one yields another minimizer, and adding or multiplying 0/1-solutions produces new solutions, enabling compositional construction of optimizers.
  • The effective dynamics live in a space of dimension at most the number of monomial potentials, so training exhibits progressive dimension reduction with finitely many eigenvalue crossings, interpreting phase-transition behavior such as grokking.
  • For group-invariant target functions, the sample complexity scales as $1/(|G|\epsilon^{1+1/\theta(\kappa+1)})$ for finite $G$ and $\mathrm{vol}(M^d/G)/\epsilon^{1+1/\theta(\kappa+1)}$ for continuous groups, so symmetry directly reduces data requirements.
  • The framework applies to any reasoning task whose loss depends on parameters through monomials, not just Abelian group addition, so the same analysis carries to broader algebraic reasoning settings.

Reading between the lines

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

  • The authors defer a complex extension in Appendix A.1; until one is supplied, Theorem 4.2 is not formally valid for the paper's own Abelian-group example, whose Fourier coefficients are complex. A rigorous complex Hermite-based proof would close the gap.
  • Because $\mathrm{O}(d)$-equivariance is assumed rather than demonstrated for the toy architecture, a direct experimental check of the off-diagonal Gram entries on that architecture would sharply test whether this is a benign assumption or a real constraint.
  • One testable extension: if the decoupling picture is correct, architectures with imposed group invariance should show earlier and sharper grokking-like transitions than unconstrained architectures on the same symbolic task, a prediction that does not follow from the paper's theorems but is a natural corollary of its mechanism.
  • The measure-space semi-ring suggests an inverse problem: given weights for a composed task and weights for one subtask, recover the unseen subtask via algebraic factorization in the potential space, something the paper mentions as future direction but does not develop.
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

4 major / 5 minor

Summary. The paper proposes a measure-theoretic framework to explain the emergence of discrete symbolic structure during continuous neural network training. It lifts neural parameters to probability measures, models training as Wasserstein gradient flow, and claims that under Gaussian initialization, odd-degree monomial potentials, and an O(d)-equivariant velocity field, the dynamics of monomial potentials decouple into independent coordinate-wise gradient flows (Theorem 4.2). It further claims a progressive reduction in the effective degrees of freedom (Theorems 4.3 and 4.4), an algebraic semi-ring structure on the space of measures (Theorem 3.3), and sample-complexity guarantees for invariant learning (Theorem 5.1). The Abelian group composition task of Section 2.1 is used throughout as the motivating example.

Significance. If the main decoupling theorem were correct, the framework would offer a novel and potentially influential explanation of how continuous gradient-based training can converge to discrete symbolic solutions, complementing the finite-neuron analysis of Tian [2024] with a measure-theoretic perspective. The paper is ambitious, clearly written in places, and includes detailed appendix proofs, an explicit acknowledgement of idealizations, and falsifiable sample-complexity predictions. The algebraic semi-ring construction (Theorem 3.3) is a genuine and correctly argued contribution. However, the central load-bearing result is not established: the assumptions of Theorem 4.2 are internally inconsistent, the motivating complex-valued example is explicitly excluded by the proof assumptions, and the degree-of-freedom reduction results are largely forced by construction rather than derived from the training dynamics. These issues undermine the paper's main claims.

major comments (4)
  1. [Assumption 4.1 / Theorem 4.2] Assumption 4.1(2) and (3) are mutually inconsistent for R=-I. By Lemma A.12, the velocity field is V_t(z)=Σ_i ∂_{ρ_i}L(ρ_t)∇r_i(z). Under Assumption 4.1(2), each r_i is an odd homogeneous polynomial of degree at least 3, so each ∇r_i is even, hence V_t(-z)=V_t(z). O(d)-equivariance with R=-I requires V_t(-z)=-V_t(z), forcing V_t≡0. Then Eq. (6) gives ∂_t ρ_{r_i}=0 despite C_i(t)>0, contradicting the exponential convergence ρ_{kkk}(t)=1-exp(-C_{kkk}t) in Section 4.1, where ∂_{ρ_{kkk}}L=-2 at initialization. Thus Theorem 4.2 is vacuous as stated and cannot apply to the paper's own examples.
  2. [Appendix A.1 / Section 2.1] The real-valued parameter restriction excludes the motivating example. Appendix A.1 states that all parameters are assumed real-valued and that a rigorous complex extension is left open, while the Abelian group task of Section 2.1 is formulated with complex Fourier coefficients z_{akj}, z_{bkj}, z_{ckj} and complex basis functions F_k. Since Theorem 4.2 and its proof rely on real Hermite polynomial identities, the central decoupling theorem is not established for the Abelian-group loss of Proposition 2.1. This is a load-bearing gap, not a cosmetic one.
  3. [Section 4.1 / Assumption 4.1(3)] O(d)-equivariance is not demonstrated for the constructed architecture. The authors note in Section 4.1 that Assumption 4.1(3) is 'not satisfied by default' and 'can be enforced through an appropriate choice of architecture', but no architecture or loss satisfying it is specified for the Abelian-group example. Proposition 2.1's loss is not shown to be O(d)-invariant, and no equivariant velocity field construction is given. Consequently Theorem 4.2 has no verified instantiation among the paper's examples.
  4. [Section 4.2 / Theorems 4.3 and 4.4] The degree-of-freedom reduction is forced by construction rather than derived from dynamics. The first statement of Theorem 4.4 follows immediately from Lemma A.13 because L_t is a finite-rank operator with rank at most m; it does not reflect a dynamical contraction. Theorem 4.3 characterizes entropy minimizers as an exponential family with m parameters, which is a standard consequence of the KKT conditions, but the paper does not prove that the gradient flow converges to this family or that the trajectory's effective dimension decreases. The monotone eigenvalue claim rests on the unverified condition [∇³L]_i∇_i L ⪰ 0, which is not checked for the Abelian loss, and the assertion of finitely many zero crossings is not rigorously derived from the non-increasing property. Thus the 'progressive contraction on the degree of freedom' advertised in the abstract is not established as a dynamical phenomenon.
minor comments (5)
  1. [Section 4.1, after Eq. (6)] The displayed formula ρ_{kkk}(t)=1-exp(-C_{kkk}t) treats C_{kkk}(t) as constant, whereas Theorem 4.2 allows a time-dependent factor; the text should clarify that this is a schematic illustration.
  2. [Proposition 3.4] The statement uses z1 and z2 in place of the measures µ1 and µ2, and it numbers two statements both as '(1)'; this should be corrected for readability.
  3. [Theorem 4.4] The notation 'vi ⊂ span(R)' should be 'v_i ∈ span(R)'.
  4. [Appendix A.1] The shorthand that writes the L2(µ) inner product as L(µ) conflicts with the loss functional L(ρ); using a different symbol would avoid confusion.
  5. [Acknowledgements] 'DAPRA' should be 'DARPA'.

Circularity Check

1 steps flagged · score 4.0 of 10

Degree-of-freedom "reduction" restates the separable-loss definition; the central decoupling theorem is not circular.

  1. self definitional [Sec. 4.2, Theorem 4.4; Eq. (3) and Lemma A.13]
    "Consider loss functional in Eq. 3, all eigenfunctions corresponding to non-zero eigenvalues of second variation L(t) lie in a subspace spanned by the monomial set R, i.e, vi ⊂ span(R)... This implies that the intrinsic degrees of freedom in the system are bounded above by m."

    Eq. (3) defines H[mu]=L(rho_{r1}[mu],...,rho_{rm}[mu]), i.e. the loss is already a function of exactly m monomial-potential coordinates. Lemma A.13 shows the second variation is a finite sum over i,j of products of integrals against the same r_i, so its range is contained in span(R) and its rank is at most m. The theorem's claim that the system has at most m intrinsic degrees of freedom is therefore an unpacking of the definition of H, not a phenomenon emerging from Wasserstein dynamics; the same bound holds for any functional of m moments. Presenting this as a discovered 'degree of freedom reduction' makes the conclusion identical to the input by construction.

full rationale

The main decoupling theorem (Theorem 4.2) is not circular: it is proved from the stated initialization, degree, and symmetry assumptions via the chain-rule form of the velocity field (Lemma A.12) and the vanishing of off-diagonal Gram integrals under a symmetric measure (Lemmas A.18-A.19). The loss-reformulation Proposition 2.1 invokes the external Tian [2024] theorem, not a self-citation. Self-citations (e.g., Zhu et al. 2025) are illustrative design references and are not load-bearing. The flagged circularity is confined to the dimension-reduction narrative: the first part of Theorem 4.4 is a direct consequence of the separable form H[mu]=L(rho_{r1},...,rho_{rm}); it is a definitional restatement rather than an emergent result. Theorem 4.3 similarly relies on the standard exponential-family fact for moment-constrained entropy minimization and does not involve the gradient-flow analysis. The real-valued restriction (Appendix A.1) and the parity compatibility of Assumptions 4.1(2)-(3) are correctness/scope concerns, not circularity, and do not affect this verdict.

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

The framework is a self-contained mathematical construction under explicit assumptions, but the central emergence claim depends on several strong, non-default conditions: Gaussian initialization, odd-degree monomials, O(d)-equivariant velocity field, and real-valued parameters. The sample-complexity theorem is directly imported from prior work, and the degree-of-freedom results are forced by the finite-dimensional dependence of the loss on monomial potentials. No free parameters are fitted to data, and no new physical entities are introduced.

assumptions (7)
  • domain assumption Initial measure µ0 = N(0,I) (Assumption 4.1(1)).
    Required for the Hermite orthogonality argument that makes the Gram matrix G_ij diagonal at initialization and keeps it diagonal over time.
  • domain assumption Every monomial potential has degree at least 3 and odd degree (Assumption 4.1(2)).
    Needed for parity arguments in Lemma A.19 and Lemma A.30; it excludes many realistic losses, including degree-1 and even-degree terms.
  • ad hoc to paper The velocity field ∇ δH/δµ is O(d)-equivariant (Assumption 4.1(3); odd symmetry in Assumption A.11).
    This is the key geometric condition that forces decoupling of monomial potentials. It is not a generic property of neural training and is not verified for the Abelian group example.
  • ad hoc to paper All parameters are real-valued (Appendix A.1).
    The proofs rely on real Hermite polynomials, but the motivating Fourier parameter space is complex; the authors call this 'tacit but critical' and defer a rigorous extension.
  • ad hoc to paper The condition [∇³L(ρ_t)]_i ∇_i L ⪰ 0 for every i (Theorem 4.4).
    Needed for the claim that Hessian eigenvalues are non-increasing over time; no constructive realization is given for when this holds.
  • standard math Exponential-family form of entropy minimizers with KKT (Theorem 4.3).
    Standard calculus of variations result used to bound the dimension of entropy-minimizing measures under moment constraints.
  • standard math Frechet derivability and regularity of the loss functional and measures, including smooth compactly supported velocity fields (Appendix A.1 item 3).
    Ensures integration by parts and the Wasserstein gradient flow equation are well defined.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Why Neural Network Can Discover Symbolic Structures with Gradient-based Training: An Algebraic and Geometric Foundation for Neurosymbolic Reasoning." pith.science (2026). https://pith.science/paper/7PVWIKSY

@misc{pith2026250621797,
  author       = {Pith},
  title        = {Pith review of: Why Neural Network Can Discover Symbolic Structures with Gradient-based Training: An Algebraic and Geometric Foundation for Neurosymbolic Reasoning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7PVWIKSY}},
  note         = {Machine review of arXiv:2506.21797}
}
abstract

We develop a theoretical framework that explains how discrete symbolic structures can emerge naturally from continuous neural network training dynamics. By lifting neural parameters to a measure space and modeling training as Wasserstein gradient flow, we show that under geometric constraints, such as group invariance, the parameter measure $\mu_t$ undergoes two concurrent phenomena: (1) a decoupling of the gradient flow into independent optimization trajectories over some potential functions, and (2) a progressive contraction on the degree of freedom. These potentials encode algebraic constraints relevant to the task and act as ring homomorphisms under a commutative semi-ring structure on the measure space. As training progresses, the network transitions from a high-dimensional exploration to compositional representations that comply with algebraic operations and exhibit a lower degree of freedom. We further establish data scaling laws for realizing symbolic tasks, linking representational capacity to the group invariance that facilitates symbolic solutions. This framework charts a principled foundation for understanding and designing neurosymbolic systems that integrate continuous learning with discrete algebraic reasoning.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

29 extracted references · 15 canonical work pages

  1. [1]

    Neurosymbolic programming

    Swarat Chaudhuri, Kevin Ellis, Oleksandr Polozov, Rishabh Singh, Armando Solar-Lezama, Yisong Yue, et al. Neurosymbolic programming. Foundations and Trends in Programming Languages , 7 0 (3): 0 158--243, 2021

    work page 2021

  2. [2]

    Neurosymbolic ai: The 3 rd wave

    Artur d’Avila Garcez and Luis C Lamb. Neurosymbolic ai: The 3 rd wave. Artificial Intelligence Review, 56 0 (11): 0 12387--12406, 2023

    2023

  3. [3]

    On the paradox of learning to reason from data

    Honghua Zhang, Liunian Harold Li, Tao Meng, Kai-Wei Chang, and Guy Van Den Broeck. On the paradox of learning to reason from data. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, pages 3365--3373, 2023

    work page 2023

  4. [4]

    On the planning abilities of large language models-a critical investigation

    Karthik Valmeekam, Matthew Marquez, Sarath Sreedharan, and Subbarao Kambhampati. On the planning abilities of large language models-a critical investigation. Advances in Neural Information Processing Systems, 36: 0 75993--76005, 2023

    work page 2023

  5. [5]

    Composing global optimizers to reasoning tasks via algebraic objects in neural nets

    Yuandong Tian. Composing global optimizers to reasoning tasks via algebraic objects in neural nets. arXiv preprint arXiv:2410.01779, 2024

    arXiv 2024

  6. [6]

    Grokking: Generalization beyond overfitting on small algorithmic datasets

    Alethea Power, Yuri Burda, Harri Edwards, Igor Babuschkin, and Vedant Misra. Grokking: Generalization beyond overfitting on small algorithmic datasets. arXiv preprint arXiv:2201.02177, 2022

    arXiv 2022

  7. [7]

    Grokking modular arithmetic

    Andrey Gromov. Grokking modular arithmetic. arXiv preprint arXiv:2301.02679, 2023

    arXiv 2023

  8. [8]

    Progress measures for grokking via mechanistic interpretability, 2023

    Neel Nanda, Lawrence Chan, Tom Lieberum, Jess Smith, and Jacob Steinhardt. Progress measures for grokking via mechanistic interpretability, 2023. URL https://arxiv. org/abs/2301.05217, 2023

    arXiv 2023

Show all 29 references
  1. [9]

    On the power of over-parametrization in neural networks with quadratic activation

    Simon Du and Jason Lee. On the power of over-parametrization in neural networks with quadratic activation. In International conference on machine learning, pages 1329--1338. PMLR, 2018

    2018

  2. [10]

    Relu ^2 wins: Discovering efficient activation functions for sparse llms

    Zhengyan Zhang, Yixin Song, Guanghui Yu, Xu Han, Yankai Lin, Chaojun Xiao, Chenyang Song, Zhiyuan Liu, Zeyu Mi, and Maosong Sun. Relu ^2 wins: Discovering efficient activation functions for sparse llms. arXiv preprint arXiv:2402.03804, 2024

    2024 arXiv

  3. [11]

    Glu variants improve transformer

    Noam Shazeer. Glu variants improve transformer. arXiv preprint arXiv:2002.05202, 2020

    2002 arXiv

  4. [12]

    Searching for efficient transformers for language modeling

    David So, Wojciech Ma \'n ke, Hanxiao Liu, Zihang Dai, Noam Shazeer, and Quoc V Le. Searching for efficient transformers for language modeling. Advances in neural information processing systems, 34: 0 6010--6022, 2021

    2021

  5. [13]

    Mean-field theory of two-layers neural networks: dimension-free bounds and kernel limit

    Song Mei, Theodor Misiakiewicz, and Andrea Montanari. Mean-field theory of two-layers neural networks: dimension-free bounds and kernel limit. In Conference on learning theory, pages 2388--2464. PMLR, 2019

    2019

  6. [14]

    Optimal transport: old and new, volume 338

    C \'e dric Villani et al. Optimal transport: old and new, volume 338. Springer, 2009

    2009

  7. [15]

    Gradient flows: in metric spaces and in the space of probability measures

    Luigi Ambrosio, Nicola Gigli, and Giuseppe Savar \'e . Gradient flows: in metric spaces and in the space of probability measures. Springer Science & Business Media, 2008

    2008

  8. [16]

    Lectures on phase transitions and the renormalization group

    Nigel Goldenfeld. Lectures on phase transitions and the renormalization group. CRC Press, 2018

    2018

  9. [17]

    The exact sample complexity gain from invariances for kernel regression

    Behrooz Tahmasebi and Stefanie Jegelka. The exact sample complexity gain from invariances for kernel regression. Advances in Neural Information Processing Systems, 36, 2023

    2023

  10. [18]

    Rethinking addressing in language models via contexualized equivariant positional encoding

    Jiajun Zhu, Peihao Wang, Ruisi Cai, Jason D Lee, Pan Li, and Zhangyang Wang. Rethinking addressing in language models via contexualized equivariant positional encoding. arXiv preprint arXiv:2501.00712, 2025

    2025 arXiv

  11. [19]

    The parallelism tradeoff: Limitations of log-precision transformers

    William Merrill and Ashish Sabharwal. The parallelism tradeoff: Limitations of log-precision transformers. Transactions of the Association for Computational Linguistics, 11: 0 531--545, 2023

    2023

  12. [20]

    The illusion of state in state-space models

    William Merrill, Jackson Petty, and Ashish Sabharwal. The illusion of state in state-space models. arXiv preprint arXiv:2404.08819, 2024

    2024 arXiv

  13. [21]

    Methods of information geometry, volume 191

    Shun-ichi Amari and Hiroshi Nagaoka. Methods of information geometry, volume 191. American Mathematical Soc., 2000

    2000

  14. [22]

    Stability and generalization

    Olivier Bousquet and Andr \'e Elisseeff. Stability and generalization. Journal of machine learning research, 2 0 (Mar): 0 499--526, 2002

    2002

  15. [23]

    Understanding machine learning: From theory to algorithms

    Shai Shalev-Shwartz and Shai Ben-David. Understanding machine learning: From theory to algorithms. Cambridge university press, 2014

    2014

  16. [24]

    Equivariant architectures for learning in deep weight spaces

    Aviv Navon, Aviv Shamsian, Idan Achituve, Ethan Fetaya, Gal Chechik, and Haggai Maron. Equivariant architectures for learning in deep weight spaces. In International Conference on Machine Learning, pages 25790--25816. PMLR, 2023

    2023

  17. [25]

    Signal processing for implicit neural representations

    Dejia Xu, Peihao Wang, Yifan Jiang, Zhiwen Fan, and Zhangyang Wang. Signal processing for implicit neural representations. Advances in Neural Information Processing Systems, 35: 0 13404--13418, 2022

    2022

  18. [26]

    Model-glue: Democratized llm scaling for a large model zoo in the wild

    Xinyu Zhao, Guoheng Sun, Ruisi Cai, Yukun Zhou, Pingzhi Li, Peihao Wang, Bowen Tan, Yexiao He, Li Chen, Yi Liang, et al. Model-glue: Democratized llm scaling for a large model zoo in the wild. arXiv preprint arXiv:2410.05357, 2024

    arXiv

  19. [27]

    Self-consuming generative models go mad

    Sina Alemohammad, Josue Casco-Rodriguez, Lorenzo Luzi, Ahmed Imtiaz Humayun, Hossein Babaei, Daniel LeJeune, Ali Siahkoohi, and Richard Baraniuk. Self-consuming generative models go mad. In The Twelfth International Conference on Learning Representations, 2024

    2024

  20. [28]

    Polynomial width is sufficient for set representation with high-dimensional features

    Peihao Wang, Shenghao Yang, Shu Li, Zhangyang Wang, and Pan Li. Polynomial width is sufficient for set representation with high-dimensional features. arXiv preprint arXiv:2307.04001, 2023

    2023 arXiv

  21. [29]

    Low-dimensional invariant embeddings for universal geometric learning

    Nadav Dym and Steven J Gortler. Low-dimensional invariant embeddings for universal geometric learning. Foundations of Computational Mathematics, pages 1--41, 2024

    2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.