Pith. sign in

REVIEW 2 major objections 5 minor 26 references

A Group Theoretic Analysis of the Symmetries Underlying Base Addition and Their Learnability by Neural Networks

T0 review · 2 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 A tiny recurrent network can learn to add arbitrarily long numbers when the carry function is chosen and presented to expose its symmetry.

desk verdict A real empirical bridge between carry-function structure and RNN learnability, but the headline correlations rest on a depth-4 truncation that needs a convergence check. read the letter →

arxiv 2507.10678 v2 pith:E5ASE4OU submitted 2025-07-14 cs.LG cs.AIcs.NEq-bio.NC

Cutter Dawes , Simon Segert , Kamesh Krishnamurthy , Jonathan D. Cohen This is my paper
classification cs.LGcs.AIcs.NEq-bio.NC MSC 20J0611A6368T0728A80
keywords baseadditioncarryfunctiongroupcohomologysymmetrylearningradicalgeneralizationfractaldimensionk-equivariancerecurrentneuralnetworks
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 tries to show that base addition is a symmetry-discovery problem whose difficulty for neural networks is set by the structure of the carry function, not by the size of the base or the number of digits. Using group cohomology, it constructs every valid carry function for a base as a cocycle, and shows that some carry functions are equivalent reorderings of the same addition while others differ in how fully they preserve associativity across digit depths. It measures each carry function's fractal dimension, carry frequency, and associativity fraction, then trains a one-layer recurrent network on three-digit pairs in an interleaved digit-by-digit format. The result is that simple carry functions generalize radically to six- and ten-digit numbers, and that test accuracy across carry functions is strongly rank-correlated with all three structural measures. If correct, this gives a normative account of why humans use the standard 1-carry algorithm and a recipe for making arithmetic learnable by small networks.

What carries the argument

The load-bearing object is the carry function $f$ on pairs of digits in $Z_b$, viewed as a 2-cocycle in the group extension from $Z_b$ to $Z_{b^2}$; the k-digit analogue is the depth-$k$ carry table $F_k$, whose entry says what is carried from digit $k$ to digit $k+1$ for every pair of $k$-digit numbers. Choosing a different section of the extension changes $f$ by a coboundary, which yields the alternative carry functions; a carry is $k$-equivariant when the resulting $k$-digit addition remains associative, meaning the cocycle condition survives to that depth. The paper's structural measures are computed on these tables: box-counting dimension of the border between distinct carried values, frequency of non-zero carries, and associativity fraction over sampled triplets. For training, the equivalent machinery is the interleaved problem format, which presents digits least-to-most significant so the recurrent network must store and carry a scalar value across time steps, making the carry symmetry accessible through weight sharing over time.

What would settle it

Recompute the three structural measures at depths five, six, or higher; if their rank order across carry functions changes substantially, or if the Spearman correlations with maximum test accuracy on 6-digit (or longer) numbers drop toward zero or reverse, the paper's claim that these statistics capture learning-relevant complexity would be falsified.

Watch

Extended reading notes

Core claim

The central claim is that even a tiny neural network can achieve radical generalization in addition, provided the input format exposes the per-place carry operation and the carry function itself has the right symmetry structure. Concretely, the paper argues that a carry function for base $b$ is a normalized 2-cocycle of $Z_b$ with coefficients in $Z_b$; two carry functions are equivalent when they differ by a coboundary, so the set of valid carries for two-digit addition has size $b^{b-2}$, of which $\varphi(b)$ are Single Value carries (carrying a fixed unit) and the rest are Multiple Value carries. The paper further claims that these classes separate by structure: at depth four, Single Value carry tables have box-counting dimension near 1, Multiple Value tables stay above 1.25, and a subset of Low Dimensional Multiple Value tables keep full associativity to depth 4 or 5. Training a one-layer GRU and LSTM on all 3-digit pairs with an interleaved least-to-most-significant format, the paper reports near-perfect accuracy for Single Value carries on 6-digit numbers and high accuracy through 10 digits, with maximum 6-digit test accuracy across bases 3 to 5 ranked by fractal dimension (Spearman -0.872), carry frequency (-0.656), and associativity fraction (0.887). The discovery is the alignment between cohomological complexity measures and learnability.

Load-bearing premise

The structural measures are computed on carry tables only four digits deep, and the paper assumes that depth-4 values rank the carry functions in the same order as the true complexity a network faces when generalizing from three to six or ten digits.

Editorial extensions

If this is right

  • A one-layer GRU trained only on 3-digit pairs can add 6-digit numbers almost perfectly and 10-digit numbers with high accuracy when the carry is Single Value, showing that radical generalization does not require a large model.
  • Across bases 3 to 5, maximum 6-digit test accuracy is rank-correlated with carry-table structure: fractal dimension and carry frequency hurt, associativity helps.
  • Because the standard 1-carry is the simplest Single Value carry and aligns with ordinal digit order, the paper explains the universal human choice of carry algorithm as a learnability optimum.
  • The interleaved digit-by-digit input format is a critical inductive bias: it reduces addition to modular addition plus a remembered carry, and without it the task requires external memory or ordering machinery.

Reading between the lines

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

  • A testable extension: if the depth-4 measures are proxies for asymptotic structure, then training curricula that expose the network to progressively longer numbers should yield the same rank ordering of carry functions, whereas a curriculum that fixes depth 3 may not.
  • The group cohomology picture likely generalizes to other recursive symbolic operations, such as multiplication by a fixed modulus or carry in mixed-radix systems, where the load-bearing structure is again a cocycle condition; learning difficulty there may also be predictable from the same structural measures.
  • The results suggest one reason large language models struggle with arithmetic in standard notation: the format hides the carry symmetry behind token ordering, and an explicit scratchpad or right-aligned interleaved encoding might give even small models algorithmic addition.
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

2 major / 5 minor

Summary. The paper formalizes base addition through group cohomology, treating a carry function as a 2-cocycle and studying the resulting group extensions of Z_b by Z_b. It defines Single Value and Multiple Value carry functions, introduces depth-k carry tables, and characterizes carry functions by three quantitative measures: the box-counting dimension of the table border, the frequency of carrying, and the associativity fraction, computed at depths 1-4 for bases 3-5. It then trains a 1-layer GRU (and an LSTM, in the appendix) on 3-digit addition in an interleaved digit-pair format, with one-hot or Gaussian-smoothed digit embeddings, for all carry functions in bases 3-5, and tests on 6-10 digit numbers. The main results are that Single Value carry functions are learned most effectively and generalize far out of distribution, and that maximum 6-digit test accuracy is rank-correlated with the structural measures (Spearman rho = -0.872, -0.656, 0.887 across bases 3-5). The discussion connects these findings to human arithmetic and to the universality of the 1-carry function.

Significance. If the result holds, the paper provides a clean demonstration that a very small recurrent network can discover a carry symmetry and generalize radically when the input format is interleaved and the carry function has sufficient structure. The group-theoretic framing is elegant, the experimental design is transparent, the code is openly available, and the structural measures are computed independently of the network, so the reported correlations are genuine empirical findings rather than constructions. The main limitations are that the quantitative measures are shown to separate the carry-function classes only at depth 4, and that the correlational analysis pools across bases that differ systematically in training-example count; both issues are fixable and do not undermine the core demonstration of radical generalization for Single Value carries.

major comments (2)
  1. [§3.2–3.4, Fig. 3; §4.2.3, Fig. 7] The correlational analysis uses the depth-4 values of fractal dimension, carry frequency, and associativity fraction as the operative complexity for learning to generalize from 3-digit training to 6-digit testing. For carry functions that are only finitely equivariant—and Section 5.1 explicitly leaves open which Multiple Value functions are ∞-equivariant—iterating Equation 2 to depth 4 is not a sample of a convergent asymptotic object. The paper provides no evidence that the rank ordering of carry functions by these measures is stable across depths; if the depth-5 or depth-6 ranking differs, the reported Spearman correlations (-0.872, -0.656, 0.887) may be an artifact of the chosen truncation. The authors should report the three measures at deeper depths (or at least demonstrate rank-order stability across k = 1 to 6 for the available bases) and recompute the correlations at each depth, ideally including the depths actually tested by the network.
  2. [§4.2.2–4.2.3, Fig. 6–7] The headline correlations pool all carry functions from bases 3, 4, and 5 into a single Spearman rank correlation. Section 4.2.2 (Fig. 6) shows that test accuracy increases systematically with base, and Section 5.2.2 attributes this at least in part to the larger number of training examples at larger bases. Since base may influence both the structural measures and learning, the pooled correlations could be inflated by a base-level confound. The authors should report the correlations separately within each base, or compute partial Spearman correlations controlling for base, and should separate between-class from within-class contributions to the correlations.
minor comments (5)
  1. [§3.4] The associativity fraction is said to be based on sampled triplets, but no sample size, number of trials, or confidence intervals are reported; this matters for the precision of claims such as an associativity fraction of exactly 1 up to depth 4.
  2. [§3.2] The border definition as entries with different values from their left and upper neighbors does not state how the first row and first column are treated; please clarify the boundary convention.
  3. [§4.2.1] The Gaussian convolution envelope used for semantic embeddings is described by example only; the standard deviation, normalization, and circular padding details should be specified for reproducibility.
  4. [Appendix C.2] Roughly 15% of learning curves could not be fit by the sigmoid and are excluded; the paper should state whether there is a systematic pattern among the excluded curves (for example, concentration in the Other Multiple Value class), since this could bias the robustness analysis.
  5. [§2.4.2 and Fig. 3] The Low Dimensional Multiple Value category is defined post hoc from the same quantitative measures used in Figure 7, so the visual separation into three groups in Figure 7 is partly by construction; the paper should state this explicitly and emphasize that the non-circular evidence is the continuous correlation with learning and the generalization gap in Figure 6.

Circularity Check

1 steps flagged · score 1.0 of 10

No significant circularity: structural measures are computed independently of network training; only the Low-Dimensional-Multiple-Value label is mildly self-referential, and the headline correlations do not depend on it.

  1. self definitional [Section 2.4.2 (definition); Section 4.2.2 and Figure 7 (same groupings)]
    "In Section 3) we observe that there is a subset of Multiple Value carry functions that are notably lower in complexity than the remainder. We refer to these as Low Dimensional Multiple Value carry functions, that we characterize in Section 3)"

    The category 'Low Dimensional Multiple Value' is introduced as the subset singled out by the same structural measures (fractal dimension, carry frequency, associativity) that Section 4 later reports as separating the learning curves. So the qualitative observation that learning separates 'into the same three groupings' is partly guided by the way the labels were assigned from those measures. However, the central quantitative claim is the Spearman correlation between continuous measure values and maximum test accuracy across all carry functions, which does not rely on the category labels and is therefore an independent empirical finding.

full rationale

The paper's derivation chain is largely self-contained and not circular. The group-cohomological construction of carry functions is derived in Appendix A.2 from standard results (Isaksen 2002; Brown 2012), and the three structural measures are computed directly from carry tables at depths 1-4, with no input from the neural network training or test results. The networks are trained only on digit-pair inputs and carry-specific answer sequences; they never receive the fractal dimension, carry frequency, or associativity fraction. Thus the strong correlations in Section 4.2.3 (-0.872, -0.656, 0.887) are genuine empirical relationships rather than quantities forced by construction. The self-citations (Segert 2024; Segert & Cohen 2022) are not load-bearing: the mathematical extension is re-derived in the appendix, and the network simulations are new and independently coded. The only mild self-referential element is the 'Low Dimensional Multiple Value' label being defined by thresholds on the same measures that later appear in the learning-curve groupings; that labeling is descriptive and does not drive the headline correlations. The depth-4 truncation concern raised in the reader's take is a legitimate external-validity risk about whether the measures have converged, but it is not circularity: even if depth-4 rankings were unstable, the correlations would be a real property of the depth-4 measures, not a construction artifact. Overall, the central claim has independent empirical content, so the circularity score is low.

Assumptions & free parameters 4 free parameters · 6 assumptions · 0 invented entities

The central claim rests on standard group cohomology (axioms in the list), the architectural assumption that a small recurrent net can implement carry recursion, and the ad hoc choice of depth-4 truncated structural measures and hand-set thresholds for the 'Low Dimensional Multiple Value' class. No new physical entities are introduced; the only invented construct is the Low Dimensional Multiple Value function class, which is an empirical category rather than a postulated entity. Free parameters are mostly hyperparameters and analysis choices: the Gaussian envelope for semantic embeddings, the truncation depth k=4, the class thresholds, and the sigmoid-fit parameters.

free parameters (4)
  • Gaussian convolution envelope for semantic embeddings = e.g., digit 1 in base 5 becomes (0.2, 0.5, 0.2, 0.05, 0.05)
    Chosen by hand to embed ordinal structure into one-hot vectors; no sensitivity analysis or principled derivation, and it affects the semantic-embedding learning results in Section 4.2.1.
  • Truncation depth for fractal dimension estimation = k = 1 to 4
    The box-counting dimension is estimated only up to depth 4 (Section 3.2, Figure 3A); the paper treats the depth-4 value as the structural complexity measure for the correlation in Figure 7, without demonstrating convergence for all carry-function classes.
  • Thresholds defining Low Dimensional Multiple Value class = fractal dimension in (1.25, 1.5), associativity fraction 1 up to depth 4, lower carry frequency
    These boundaries are chosen post hoc from the data in Figure 3 to separate the observed clusters; they are descriptive rather than derived and are used in the narrative classification.
  • Sigmoid fit parameters (upper asymptote, growth rate, critical point) = not tabulated per function
    Used in the robustness analysis (Appendix C.2); fits are performed with scipy.optimize.curve_fit, and about 15 percent of curves are excluded because the fit failed, so the fitted parameters are a conditional summary rather than a direct measurement.
assumptions (6)
  • standard math Group axioms and existence of finite cyclic groups Zb
    Basis for the cohomological construction in Section 2.1 and Appendix A.2.
  • domain assumption Isaksen's 2-digit cohomological construction of base addition is correct and extends iteratively to arbitrary digit length
    Section 2.3 and Appendix A.2 rely on Isaksen (2002) and Segert (2024) for the claim that carrying via a cocycle yields a faithful base representation; the paper reproduces the derivation but inherits any gaps in those sources.
  • domain assumption The interleaved (column-wise, right-to-left) training format exposes the relevant symmetry and is an appropriate model of human learning
    Section 4.1.3 states this format is used because it matches how humans learn column addition, and the radical-generalization results in Section 4.2.2 depend on it; no comparison to the standard whole-number format is made.
  • domain assumption A one-layer GRU/LSTM with hidden dimension b can implement and learn the carry recursion given enough training
    Section 4.1.2 assumes this architecture has the inductive bias (weight sharing over time) needed to discover the carry function; the paper does not prove what class of functions such a network can represent.
  • ad hoc to paper Depth-4 estimates of the quantitative measures are representative of the asymptotic structure of carry functions
    Used to build Figure 3 and the correlations in Figure 7; the paper provides no convergence argument and acknowledges at least 5-equivariance only for Low Dimensional Multiple Value carries.
  • ad hoc to paper Box-counting dimension of the carry-table border is the right measure of algorithmic complexity for learning
    Section 3.2 chooses the border (entries differing from left/upper neighbors) and minimizes over digit orderings; this is motivated by the self-similarity of carry tables but not derived from learning theory.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Group Theoretic Analysis of the Symmetries Underlying Base Addition and Their Learnability by Neural Networks." pith.science (2026). https://pith.science/paper/E5ASE4OU

@misc{pith2026250710678,
  author       = {Pith},
  title        = {Pith review of: A Group Theoretic Analysis of the Symmetries Underlying Base Addition and Their Learnability by Neural Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/E5ASE4OU}},
  note         = {Machine review of arXiv:2507.10678}
}
read the original abstract

A major challenge in the use of neural networks both for modeling human cognitive function and for artificial intelligence is the design of systems with the capacity to efficiently learn functions that support radical generalization. At the roots of this is the capacity to discover and implement symmetry functions. In this paper, we investigate a paradigmatic example of radical generalization through the use of symmetry: base addition. We present a group theoretic analysis of base addition, a fundamental and defining characteristic of which is the carry function -- the transfer of the remainder, when a sum exceeds the base modulus, to the next significant place. Our analysis exposes a range of alternative carry functions for a given base, and we introduce quantitative measures to characterize these. We then exploit differences in carry functions to probe the inductive biases of neural networks in symmetry learning, by training neural networks to carry out base addition using different carries, and comparing efficacy and rate of learning as a function of their structure. We find that even simple neural networks can achieve radical generalization with the right input format and carry function, and that learnability is closely correlated with carry function structure. We then discuss the relevance this has for cognitive science and machine learning.

Figures

Figures reproduced from arXiv: 2507.10678 by the authors.

Figure 1
Figure 1. The 16 carry tables in base 4. The Single Value carry functions are outlined [PITH_FULL_IMAGE:figures/full_fig_p010_1.png] view at source ↗
Figure 2
Figure 2. Structure at increasing depths of the three carry tables for [PITH_FULL_IMAGE:figures/full_fig_p011_2.png] view at source ↗
Figure 3
Figure 3. Quantitative characterization of different carry functions for bases 3 to 5 and [PITH_FULL_IMAGE:figures/full_fig_p013_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Learning for symbolic (one-hot) embeddings of digits. Performance over the [PITH_FULL_IMAGE:figures/full_fig_p016_4.png]
Figure 5
Figure 5. Figure 5: Learning for semantic embeddings of digits. Performance over the course [PITH_FULL_IMAGE:figures/full_fig_p017_5.png]
Figure 6
Figure 6. Figure 6: Out-of-domain generalization for symbolic embeddings of digits. For each [PITH_FULL_IMAGE:figures/full_fig_p018_6.png]
Figure 7
Figure 7. Figure 7: Relationship of carry function structure and learning, as measured by max [PITH_FULL_IMAGE:figures/full_fig_p019_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

26 extracted references · 17 canonical work pages

  1. [1]

    , year 1978

    author Anderson, P.W. , year 1978 . title Local moments and localized states . journal Science volume 201 , pages 307--316

    work page 1978

  2. [2]

    , author Hamin, M

    author Beukers, A. , author Hamin, M. , author Norman, K.A. , author Cohen, J.D. , year 2024 . title When working memory may be just working, not memory. journal Psychological Review volume 131 , pages 563

    work page 2024

  3. [3]

    , year 2012

    author Brown, K. , year 2012 . title Cohomology of Groups . Graduate Texts in Mathematics, publisher Springer New York

    work page 2012

  4. [4]

    , author Van Merri \"e nboer, B

    author Cho, K. , author Van Merri \"e nboer, B. , author Gulcehre, C. , author Bahdanau, D. , author Bougares, F. , author Schwenk, H. , author Bengio, Y. , year 2014 . title Learning phrase representations using rnn encoder-decoder for statistical machine translation . journal arXiv preprint arXiv:1406.1078

    arXiv 2014

  5. [5]

    , author McClelland, J.L

    author Cleeremans, A. , author McClelland, J.L. , year 1991 . title Learning the structure of event sequences. journal Journal of Experimental Psychology: General volume 120 , pages 235

    work page 1991

  6. [6]

    , year 1997

    author Dehaene, S. , year 1997 . title The Number Sense: How the Mind Creates Mathematics . publisher Oxford University Press , address New York

    work page 1997

  7. [7]

    Your Context Is Not an Array: Unveiling Random Access Limitations in Transformers

    author Ebrahimi, M. , author Panchal, S. , author Memisevic, R. , year 2024 . title Your context is not an array: Unveiling random access limitations in transformers . journal arXiv preprint arXiv:2408.05506

    work page Pith review arXiv 2024

  8. [8]

    , year 1990

    author Elman, J.L. , year 1990 . title Finding structure in time . journal Cognitive Science volume 14 , pages 179--211 . :https://doi.org/10.1207/s15516709cog1402\_1

    doi:10.1207/s15516709cog1402 1990

Show all 26 references
  1. [9]

    , year 2014

    author Falconer, K. , year 2014 . title Fractal Geometry: Mathematical Foundations and Applications . publisher John Wiley & Sons

    2014

  2. [10]

    , author Finzi, M

    author Goldblum, M. , author Finzi, M. , author Rowan, K. , author Wilson, A.G. , year 2023 . title The no free lunch theorem, kolmogorov complexity, and the role of inductive biases in machine learning . journal arXiv preprint arXiv:2304.05366

    2023 arXiv

  3. [11]

    , author Schmidhuber, J

    author Hochreiter, S. , author Schmidhuber, J. , year 1997 . title Long Short-Term Memory . journal Neural Computation volume 9 , pages 1735--1780 . :10.1162/neco.1997.9.8.1735, http://arxiv.org/abs/https://direct.mit.edu/neco/article-pdf/9/8/1735/813796/neco.1997.9.8.1735.pdf...

    1997 doi

  4. [12]

    , author Kahana, M.J

    author Howard, M.W. , author Kahana, M.J. , year 2002 . title A distributed representation of temporal context . journal Journal of mathematical psychology volume 46 , pages 269--299

    2002

  5. [13]

    , year 2002

    author Isaksen, D. , year 2002 . title A cohomological viewpoint on elementary school arithmetic . journal The American Mathematical Monthly

    2002

  6. [14]

    , author Benna, M.K

    author Ji-An, L. , author Benna, M.K. , author Mattar, M.G. , year 2024 . title Discovering cognitive strategies with tiny recurrent neural networks . journal bioRxiv https://www.biorxiv.org/content/early/2024/10/05/2023.04.12.536629, :10.1101/2023.04.12.536629, http://arxiv.o...

    2024 doi

  7. [15]

    , author Sutskever, I

    author Krizhevsky, A. , author Sutskever, I. , author Hinton, G.E. , year 2012 . title Imagenet classification with deep convolutional neural networks , in: booktitle Advances in Neural Information Processing Systems , publisher Curran Associates, Inc

    2012

  8. [16]

    , author Bottou, L

    author Lecun, Y. , author Bottou, L. , author Bengio, Y. , author Haffner, P. , year 1998 . title Gradient-based learning applied to document recognition . journal Proceedings of the IEEE volume 86 , pages 2278--2324 . :10.1109/5.726791

    1998 doi

  9. [17]

    , author Poiitis, M

    author Loukas, A. , author Poiitis, M. , author Jegelka, S. , year 2021 . title What training reveals about neural network complexity . journal Advances in Neural Information Processing Systems volume 34 , pages 494--508

    2021

  10. [18]

    , author Jiang, Z

    author Nogueira, R. , author Jiang, Z. , author Lin, J. , year 2021 . title Investigating the limitations of transformers with simple arithmetic tasks . journal arXiv preprint arXiv:2102.13019

    2021 arXiv

  11. [19]

    , year 2023

    author Piantadosi, S.T. , year 2023 . title The algorithmic origins of counting . journal Child Development volume 94 , pages 1472--1490

    2023

  12. [20]

    , author Wang, H

    author Qian, J. , author Wang, H. , author Li, Z. , author Li, S. , author Yan, X. , year 2022 . title Limitations of language models in arithmetic and symbolic induction . journal arXiv preprint arXiv:2208.05051

    2022 arXiv

  13. [21]

    , year 2024

    author Segert, S. , year 2024 . title Maximum Entropy, Symmetry, and Relational Bottleneck: Unraveling the Impact of Inductive Biases on Systematic Reasoning . Ph.D. thesis. Princeton University

    2024

  14. [22]

    , author Cohen, J

    author Segert, S. , author Cohen, J. , year 2022 . title A self-supervised framework for function learning and extrapolation . journal Transactions on Machine Learning Research https://openreview.net/forum?id=ILPFasEaHA

    2022

  15. [23]

    , author Holyoak, K.J

    author Webb, T. , author Holyoak, K.J. , author Lu, H. , year 2023 . title Emergent analogical reasoning in large language models . journal Nature Human Behavior

    2023

  16. [24]

    , year 1992

    author Wynn, K. , year 1992 . title Addition and subtraction by human infants . journal Nature volume 358 , pages 749--750

    1992

  17. [25]

    , author Campbell, D

    author Yang, Y. , author Campbell, D. , author Huang, K. , author Wang, M. , author Cohen, J. , author Webb, T. , year 2025 . title Emergent symbolic mechanisms support abstract reasoning in large language models . https://arxiv.org/abs/2502.20332, http://arxiv.org/abs/2502.20...

    2025 arXiv

  18. [26]

    , author Nova, A

    author Zhou, H. , author Nova, A. , author Larochelle, H. , author Courville, A. , author Neyshabur, B. , author Sedghi, H. , year 2022 . title Teaching algorithmic reasoning via in-context learning . journal arXiv preprint arXiv:2211.09066

    2022 arXiv

Pith tools

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