Pith. sign in

REVIEW 4 major objections 7 minor 1 cited by

Data-driven Machine Learning Cannot Reach Symbolic-level Logical Reasoning -- The Limit of the Scaling Law

T0 review · 4 major / 7 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read Data-driven machine learning cannot reach symbolic-level logical reasoning by scaling alone, because syllogistic correctness is structural and end-to-end training creates contradictory targets.

desk verdict Real, narrow empirical findings buried under an overbroad universality claim; the structural argument silently assumes traditional 24-mood syllogistic with existential import, so the universal negative doesn't hold as stated. read the letter →

arxiv 2606.26454 v2 pith:GRISKCCI submitted 2026-06-24 cs.AI

Tiansi Dong , Mateja Jamnik , Pietro Liò This is my paper
classification cs.AI
keywords syllogisticreasoningsymbolicscalinglawsupervisedlearningEulerdiagramslargelanguagemodelslogicaltrainingdatalimits
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

Data-driven machine learning cannot attain the rigour of symbolic-level logical reasoning merely by scaling up data and training time, because correctness in syllogistic reasoning is determined by logical structure, not statistical regularity. The paper identifies two architecture-independent obstacles: the training-data combination table cannot distinguish all 24 valid syllogistic forms (some premise pairs legitimately support two different conclusions under the traditional reading), and end-to-end premise-to-conclusion mapping forces the pattern-recognition component to complete partial objects—an ability that the reasoning component must suppress to avoid inventing new entities. Experiments with an image-input network (Euler Net and its extension SupEN) show accuracy rising with added data but peaking at 97.8% on random tests and 76% over all 24 valid types, while GPT-5 reaches 100% decisions on some benchmarks yet still produces wrong explanations. The conclusion is that empirical training that stops at 100% accuracy leaves no optimisation pressure to correct reasoning defects, so the scaling law cannot guarantee symbolic-level logical reasoning.

What carries the argument

The combination table is the load-bearing device for the first deficit: it maps premise-pair images to conclusion vectors, and because the four syllogistic relations are realised by overlapping Euler diagrams, the same cell can contain several valid conclusions (e.g., BARBARA and BARBARI from the same premises), so the supervised target is not unique. The second deficit is carried by the end-to-end architecture itself, specifically its Siamese pattern-recognition front end, which shares weights and is trained to complete whole images from parts; this is the behaviour that lets the network hallucinate absent circles. The paper also offers an analogy: the self-attention matrix in transformers

What would settle it

Under a semantics where 'all' does not imply existence, search all 256 premise-conclusion pairs for any case in which the same two premises yield two distinct valid conclusions. If no such pair exists, Deficit 1's contradiction arises solely from existential import; a network trained only on the 15 Boolean-valid moods should then be checked for whether it still exhibits contradictory targets. If it does not, the claim that training data cannot distinguish valid syllogisms is not a general limit of scaling.

Watch

Extended reading notes

Core claim

The paper's central claim is that no conventional supervised system, trained purely on input-output data, can be guaranteed to reach symbolic-level syllogistic reasoning, even with infinite data. Two deficits are argued to be unavoidable: (1) the combination table that generates training examples maps multiple premise types to the same cells, so identical premises can be paired with different valid conclusions—making the target ambiguous; (2) the end-to-end mapping from premise images to conclusion vectors pits pattern completion against logical deduction, because the perceptual front end may reconstruct whole objects from partial ones (e.g., half-circles becoming full circles) while the rea

Load-bearing premise

The central claim's first deficit relies on the traditional Aristotelian reading under which 'some' follows from 'all' (existential import), giving 24 valid syllogisms; under the standard Boolean reading only 15 are valid and the conflicting-target phenomenon disappears. The paper never states this dependence.

Editorial extensions

If this is right

  • A supervised system that stops when it hits 100% benchmark accuracy loses the only training signal that could correct flawed explanations, so empirical accuracy is a ceiling, not a step toward symbolic reasoning.
  • Training data cannot exhaust unintended inputs: each new corrective example (e.g., a 270-degree arc) creates new unintended patterns, so coverage-based scaling loops forever.
  • Apparent geometric competence in image-input networks may be an illusion; changing circle colour from green drops Euler Net's accuracy to 8.1%, indicating the network relies on surface features rather than abstract relations.
  • To get guaranteed symbolic-level reasoning, a network must construct and inspect an explicit model (as SphNN does with spheres), rather than learn a function from premises to conclusions.
  • The same reasoning applies to more advanced logical tasks: if single-step syllogistic reasoning is beyond the reach of scaling, then multi-step logical reasoning is even further out of reach.

Reading between the lines

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

  • The 24-mood catalogue depends on existential import ('some' from 'all'); under the standard modern Boolean reading only 15 moods are valid, and then no premise pair has two valid conclusions—so Deficit 1's contradictory-target claim is not architecture-independent but convention-dependent, a fact the paper does not flag.
  • A direct test follows: train a supervised network on only the 15 Boolean-valid syllogisms and probe whether it shows conflicting-target errors on the five subaltern moods; if it does not, the 'fundamental limitation' reflects the logical framework, not the learning paradigm.
  • The paper's colour-shift experiment suggests a broader caution for neural interpretability: latent feature vectors that look geometric to human observers may encode colour and surface statistics, so explanations built on 'what the network must have learned' can be ungrounded.
  • If the argument is correct, the burden shifts to demonstrating any learning-based system that does not explicitly build models can nonetheless guarantee validity—an open problem that the paper leaves implicit.
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 / 7 minor

Summary. The paper argues that supervised data-driven systems, even under unlimited scaling of data and training time, cannot reach the authors' standard of symbolic-level syllogistic reasoning. It identifies two purported fundamental limitations of image-input supervised learners: (i) the Euler-diagram combination table cannot separate all 24 valid Aristotelian syllogistic moods, because several moods share the same premise diagrams but have different valid conclusions, and (ii) end-to-end premise-to-conclusion mapping forces the pattern-recognition component to complete or inject objects, creating contradictory training targets. The authors support this with experiments on Euler Net and its extension SupEN: random OOD test accuracy improves from 56% to 97.8% over 19 self-training loops, yet per-mood accuracies are as low as 50%; a colour-shift control drops accuracy to 8.1%; and GPT-5 / GPT-5-nano achieve near-100% decision accuracy on syllogistic benchmarks but produce incorrect or hallucinated explanations. The paper concludes that increasing data and training time alone cannot ensure symbolic-level logical reasoning.

Significance. If the theoretical claims were rigorously established, the paper would be a significant challenge to the scaling-law narrative for logical reasoning. The empirical contributions are valuable: the per-mood evaluation of Euler Net across all 24 traditional moods, the explanation-quality metric for GPT-5, the colour-shift control, and the demonstration of an endless loop of unintended inputs are all useful data points. The paper also makes a fair point that benchmark accuracy alone is not evidence of symbolic-level competence. However, the theoretical pillars are considerably weaker than the empirical ones: the first 'fundamental' limitation depends on an unstated logical convention, the second is not formalised as an impossibility, and the softmax argument is not a valid no-go proof. The paper would need substantial reworking and a narrowed claim before its central conclusion can be accepted.

major comments (4)
  1. [Section 4.1, Figure 5, Table 3] Deficit 1 rests entirely on the 24-mood traditional syllogistic in which 'all' carries existential import. Under the standard Boolean reading, the five subaltern moods (BARBARI, CELARONT, CESARO, CAMESTROS, CALEMOS) are invalid; for each premise pair there is at most one Boolean-valid conclusion, and the alleged contradictory-target ambiguity disappears. The paper nowhere flags this reliance on existential import; it presents the 24-mood count and the resulting deficit as architecture-independent facts. This is load-bearing because Deficit 1 is one of the two 'fundamental limitations' in the abstract and Section 4. Please explicitly state the traditional convention, discuss the Boolean alternative, and either restrict all impossibility claims to the existential-import framework or provide a version of Deficit 1 that does not depend on subaltern moods. Table 1's 50% accuracy on BARBARI is
  2. [Section 5, footnote 1] The softmax argument is not a valid impossibility proof. Softmax normalises a single output distribution to sum to one, but a Transformer can output 'all' in BARBARA contexts and 'some' in BARBARI contexts because those are different inputs/tasks. The apparent conflict arises only because the same premise pair is assigned two different correct conclusions under the 24-mood convention, which is a labelling ambiguity, not a property of softmax. To make a formal claim about self-attention, the authors need to specify the input/output encoding and prove a lower bound or an information-theoretic obstruction. As written, this section does not support the conclusion that scaling cannot help.
  3. [Section 4.2 and Experiment 1 (Section 6.1)] Deficit 2 is not established as a fundamental limitation. The experiments show that a particular Siamese-MAE component, when fed out-of-distribution half-circles, completes them or ignores them; this is an empirical failure of OOD generalisation. The argument that 'new training data generates new unintended inputs' demonstrates that a finite training set cannot enumerate a continuous input space, but it does not show that no learning rule can generalise to OOD inputs. To claim 'cannot', the authors need a formal no-go result under explicit architectural and data assumptions, or they should reduce the claim to an empirical observation about the specific systems tested. As stated, the conclusion is overgeneralised.
  4. [Section 6.4, Table 2] The GPT-5 experiments demonstrate that current models can make correct decisions with wrong explanations, but they do not establish that scaling cannot improve explanation quality. The sentence '100% accuracy is the maximum performance guided by the scaling law, usually accompanied by the stop of training' assumes that training halts at decision accuracy and that explanation correctness is not part of the objective; this is an experimental design choice, not a theorem. If one trained with an objective that includes explanation correctness, the argument would not apply. Please separate the empirical observation from the theoretical conclusion.
minor comments (7)
  1. [Section 2] The definition of 'valid' should explicitly state that the paper adopts the traditional Aristotelian reading with existential import. Without this, Table 3's list of 24 valid moods is confusing to readers familiar with the Boolean 15-mood system.
  2. [Table 1] The per-mood accuracies have no sample sizes, confidence intervals, or statistical tests. Several values (50%, 66.7%, 75%) may be based on very small numbers of trials, which weakens the empirical claim that these types are not mastered.
  3. [Section 6.2] The relationship between D1 and D_T is unclear: D1 is described as both a new test data set and part of the training data. Please clarify how the same data can serve both roles without leaking into the evaluation, and justify the 9:1 ratio of D2 to D1.
  4. [Algorithm 1] The parameters Threshold, maxSize, maxTime, and the random-generation constraints (minimum radius 0.1, circles fully inside the image) are not given concrete values. For reproducibility, these should be reported.
  5. [Figure 12] The oscillation is attributed to random search, but no multiple independent runs or error bars are shown. It is unclear whether the peak at loop 19 is stable or a single-run artifact.
  6. [Section 6.3] The colour-shift result is striking, but the number of trials and variance are not reported. Adding this information would make the 8.1% figure interpretable.
  7. [Throughout] There are several typos and minor inconsistencies, e.g., 'doublele words' in Section 6.4, 'Tesing' in the Section 6.2 heading, and occasional unclear cross-references to subfigures. A careful proofread is needed.

Circularity Check

1 steps flagged · score 4.0 of 10

Deficit 1's 'contradictory targets' are built into the paper's 24-mood convention and single-vector output encoding; the broader scaling conclusion retains independent empirical content.

  1. self definitional [Section 2 (validity definition), Section 4.1 / Figure 5, Appendix Table 3]
    "If we allow two terms in premises to change positions and fix the order of terms in the conclusion statement, there will be 256 different forms of Aristotelian syllogistic reasoning, among which 24 types (listed in Table 3 in the Appendix) are valid [26] ... [Figure 2:] the output is a vector representing all possible relations between the subject and the predicate. ... This results in the fact that the combination table cannot distinguish each valid type of syllogistic reasoning, as shown in Figure 5 – the two syllogistic relations some ... are ... and some ... are not ... occupy three rows,"

    The alleged 'deficit 1' is an immediate consequence of the paper's own task definition. With the traditional 24-mood syllogistic (subaltern moods such as BARBARI included), a premise pair such as BARBARA/BARBARI has two valid conclusions; with a single 4-element output vector, these become two mutually exclusive labels for the same input. Thus 'training data cannot distinguish all 24 valid types' and the 'contradictory targets' are not an architecture-independent discovery but are built into the chosen validity convention and output code. A multi-label/24-way encoding, or the Boolean 15-mood reading, would remove the conflict. The paper presents this constructed ambiguity as a fundamental limitation of scaling.

full rationale

The Euler Net/SupEN experiments and GPT-5 tests are genuine external measurements, giving the paper independent empirical content. The OOD/unintended-input argument is a near-tautological but non-circular point that finite data cannot cover infinite input spaces. Self-citations to SphNN and Euler Net are background or objects of study rather than load-bearing proofs. The central circularity concern is limited to Deficit 1 (Section 4.1), where the 'contradictory targets' follow by construction from the chosen 24-mood validity set and the single four-way output vector; this pillar partially reduces to the authors' own design choices. The Boolean-reading objection is a correctness/scope limitation rather than a circularity, but it reinforces that the deficit is convention-dependent. Score 4 reflects partial circularity in one load-bearing pillar while the central claim retains independent content.

Assumptions & free parameters 6 free parameters · 6 assumptions · 1 invented entities

The theoretical deficits are largely parameter-free: no fitted constants drive the two structural arguments. The free parameters listed shape only the experiments (SupEN training loops, thresholds, random-generation constraints). The axioms carry the real burden: (1) the traditional 24-mood system with existential import, on which the 'contradictory targets' pillar rests; (2) the Euler-diagram semantics inherited from the authors' own SphNN/Euler Net work; (3) the collapse of three 'some' configurations into one diagram relation, a design choice of the authors' combination table; (4) the strong OOD criterion that makes the negative conclusion partially definitional; plus (5,6) the asserted inevitability claims in Sections 4.2 and 5.

free parameters (6)
  • Unintended-output target vector [0,0,0,0] = [0,0,0,0]
    Section 6.1: the all-zero output vector is a hand-chosen target class for single-circle inputs; its definition (loss above Threshold in Algorithm 1) drives the 'never-ending loop' conclusion that training data cannot exhaust unintended inputs.
  • Threshold, maxSize, maxTime (Algorithm 1)
    Algorithm 1: inputs are 'unintended' when cross-entropy loss exceeds Threshold; all three parameters are hand-set with no values or sensitivity analysis reported.
  • Random-generation constraints (minimum circle radius 0.1; circles fully inside image) = radius >= 0.1
    Section 6.2: these hand-chosen constraints define the OOD test distribution and the search space for new training data; the 56%-to-97.8% trajectory depends on them.
  • D2:D1 = 9:1 data ratio = 9:1
    Section 6.2: original training data D2 is 9 times larger than newly created D1; chosen by hand without ablation.
  • Number of improvement loops (peak reported at loop 19) = 19 loops (of 20)
    Section 6.2, Figure 12: the headline 97.8% is the selected peak of an oscillating accuracy trajectory, with the oscillation admitted to be an artifact of random search.
  • Arc-length test points for half-circle experiments = 0, 60, 120, 180, 240, 300, 360 degrees
    Section 6.1, Figure 10: the monotonic confidence-vs-arc-length relation is demonstrated at hand-chosen angles; no grid justification is given.
assumptions (6)
  • domain assumption Traditional 24-mood syllogistic with existential import: subaltern moods (BARBARI, CELARONT, CESARO, CAMESTROS, CALEMOS) are counted as valid.
    Table 3 lists 24 valid types without discussing the Boolean 15-mood alternative. The 'contradictory targets' in Deficit 1 (identical premises, two valid conclusions) exist only because of these subaltern moods.
  • domain assumption Syllogistic relations map one-to-one to Euler-diagram relations P and ¬D over non-empty sets: 'all X are Y' iff P(OX,OY); 'some X are Y' iff not D(OX,OY); 'no X are Y' iff D(OX,OY); 'some X are not Y' iff not P(OX,OY).
    Section 2 and Figure 4 caption ('non-empty sets'). Requires existence of terms; breaks under empty-set semantics. This mapping, from the authors' own SphNN line, is the representational bedrock of Deficit 1.
  • ad hoc to paper The three set-theoretic realisations of 'some X are Y' (W subset of V; V subset of W; W intersects V) collapse into the single diagram relation not D(OX,OY) in the combination table.
    Section 4.1, Figure 5: this collapsing of configurations is the source of the claimed indistinguishability; it is a design choice of the authors' own Euler Net [30,31], not a necessity of image-input learning.
  • domain assumption Symbolic-level reasoning requires perfect handling of all unintended (out-of-distribution) inputs and counterexample construction for invalid moods.
    Section 2: 'This criterion also applies to neural networks reasoning with out-of-distribution data (unintended inputs).' This strong criterion, drawn from the authors' SphNN work [19,20], makes the negative conclusion partially definitional: no finite training set can exhaust an unbounded input class.
  • ad hoc to paper Pattern-recognition components inevitably recognise the whole from its parts, creating objectives contradictory to the reasoning component.
    Section 4.2: inevitability is asserted from MAE/Siamese-MAE behaviour [45,46] and observed in SupEN; no proof is given that the conflict is unavoidable across architectures or training regimes.
  • ad hoc to paper Self-attention is a learned combination table, and softmax normalisation prevents 100% accuracy on both 'all' and 'some' conclusions in the same context.
    Section 5: the asserted equivalence between the attention matrix and the syllogism combination table, and the claim that two alternative valid conclusions must both receive probability 1, are not derived from the task's correctness condition; the two moods are alternatives, not simultaneous requirements.
invented entities (1)
  • Unintended-input class with all-zero output target [0,0,0,0]
    purpose: A synthetic output label allowing SupEN to flag single-circle and partial-circle inputs as out-of-domain (Section 6.1).
    The all-zero target vector is an ad hoc experimental construct, not a discovered category; its definition (loss-above-threshold) determines Experiment 1's conclusion that new training data generates new unintended inputs endlessly.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Data-driven Machine Learning Cannot Reach Symbolic-level Logical Reasoning -- The Limit of the Scaling Law." pith.science (2026). https://pith.science/paper/GRISKCCI

@misc{pith2026260626454,
  author       = {Pith},
  title        = {Pith review of: Data-driven Machine Learning Cannot Reach Symbolic-level Logical Reasoning -- The Limit of the Scaling Law},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GRISKCCI}},
  note         = {Machine review of arXiv:2606.26454}
}
read the original abstract

By promoting vectors to spheres and enabling explicit model construction, neural networks can perform symbolic-level syllogistic reasoning without training data. We identify two fundamental limitations that prevent conventional data-driven machine learning systems from achieving this capability: training data generated by the combination table cannot distinguish all 24 valid syllogism types, and end-to-end premise-to-conclusion mapping creates contradictory targets within neural components. Experiments with two representative conventional systems, GPT-5 using linguistic inputs and Euler Net using visual inputs, support this analysis. ChatGPT GPT-5 may reach 100% accuracy in syllogistic reasoning, but with hallucinations. Because the learning process terminates upon reaching 100% accuracy, the system cannot progress beyond empirical accuracy to symbolic level reasoning. Random test data reduced Euler Net's accuracy to 56%. Repeatedly expanding the training set increased its accuracy to 97%, with perfect performance on 8 syllogism types. However, because unintended inputs cannot be exhaustively covered, even 100% test accuracy does not imply symbolic-level reasoning. Since syllogistic reasoning underpins logical reasoning and human rationality, these results suggest that increasing data and training time alone cannot ensure symbolic level logical reasoning.

Review history (2 revisions) →

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Logical Judgments Under Pressure: Diagnosing Syllogistic Stability with Learned Soft Prefixes

    cs.AI 2026-07 conditional novelty 6.0 of 10

    Learned soft prefixes reliably flip correct syllogistic judgments in LLMs, transferring across unseen forms and interfaces and behaving mainly as a broad answer preference rather than a transferable logical operation.

Reference graph

Works this paper leans on

51 extracted references · 11 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Nature619, 686–689 (2023)

    Biever, C.: ChatGPT broke the turing test — the race is on for new ways to assess AI. Nature619, 686–689 (2023)

    2023

  2. [2]

    Nature550, 354–359 (2017)

    Silver, D., Schrittwieser, J., Simonyan, K., Antonoglou, I., Huang, A., Guez, A., Hubert, T., Baker, L., Lai, M., Bolton, A., Chen, Y., Lillicrap, T., Hui, F., Sifre, L., Driessche, G., Graepel, T., Hassabis, D.: Mastering the game of go without human knowledge. Nature550, 354–359 (2017)

    2017

  3. [3]

    Nature 588, 604–609 (2020) 17

    Schrittwieser, J., Antonoglou, I., Hubert, T., Simonyan, K., Sifre, L., Schmitt, S., Guez, A., Lockhart, E., Hassabis, D., Graepel, T., Lillicrap, T., Silver, D.: Mastering Atari, Go, chess and shogi by planning with a learned model. Nature 588, 604–609 (2020) 17

    2020

  4. [4]

    Nature (2024)

    Abramson, J., Adler, J., Dunger, J., al.: Accurate structure prediction of biomolecular interactions with alphafold 3. Nature (2024)

    2024

  5. [5]

    Nature600(7887), 70–74 (2021)

    Davies, A., Velickovic, P., Buesing, L., Blackwell, S., Zheng, D., Tomasev, N., Tanburn, R., Battaglia, P.W., Blundell, C., Juh´ asz, A., Lackenby, M., Williamson, G., Hassabis, D., Kohli, P.: Advancing mathematics by guiding human intuition with AI. Nature600(7887), 70–74 (2021)

    2021

  6. [6]

    Nature625, 476–482 (2024)

    Trinh, T.H., Wu, Y., Le, Q.V., He, H., Luong, T.: Solving olympiad geometry without human demonstrations. Nature625, 476–482 (2024)

    2024

  7. [7]

    https://arxiv.org/abs/2001.08361

    Kaplan, J., McCandlish, S., Henighan, T., Brown, T.B., Chess, B., Child, R., Gray, S., Radford, A., Wu, J., Amodei, D.: Scaling Laws for Neural Language Models (2020). https://arxiv.org/abs/2001.08361

    arXiv 2020

  8. [8]

    Proceedings of the National Academy of Sciences121(27) (2024)

    Bahri, Y., Dyer, E., Kaplan, J., Lee, J., Sharma, U.: Explaining neural scaling laws. Proceedings of the National Academy of Sciences121(27) (2024)

    2024

Show all 51 references
  1. [9]

    Creswell, A., Shanahan, M., Higgins, I.: Selection-Inference: Exploiting Large Language Models for Interpretable Logical Reasoning (2022)

    2022

  2. [10]

    Wei, J., Wang, X., Schuurmans, D., Bosma, M., Ichter, B., Xia, F., Chi, E., Le, Q., Zhou, D.: Chain-of-Thought Prompting Elicits Reasoning in Large Language Models (2023)

    2023

  3. [11]

    Lightman, H., Kosaraju, V., Burda, Y., Edwards, H., Baker, B., Lee, T., Leike, J., Schulman, J., Sutskever, I., Cobbe, K.: Let’s Verify Step by Step (2023)

    2023

  4. [12]

    In: NAACL (2024)

    Eisape, T., Tessler, M., Dasgupta, I., Sha, F., Steenkiste, S., Linzen, T.: A sys- tematic comparison of syllogistic reasoning in humans and language models. In: NAACL (2024)

    2024

  5. [13]

    PNAS Nexus3(7) (2024)

    Lampinen, A.K., Dasgupta, I., Chan, S.C.Y., Sheahan, H.R., Creswell, A., Kumaran, D., McClelland, J.L., Hill, F.: Language models, like humans, show content effects on reasoning tasks. PNAS Nexus3(7) (2024)

    2024

  6. [14]

    https://arxiv.org/abs/ 2408.08590

    Kim, G., Valentino, M., Freitas, A.: A Mechanistic Interpretation of Syllogistic Reasoning in Auto-Regressive Language Models (2025). https://arxiv.org/abs/ 2408.08590

    2025 arXiv

  7. [15]

    Lawrence Erlbaum Associates, Inc., Hove, HK, Hillsdale, USA (1991)

    Johnson-Laird, P.N., Byrne, R.M.J.: Deduction. Lawrence Erlbaum Associates, Inc., Hove, HK, Hillsdale, USA (1991)

    1991

  8. [16]

    Journal of Cognitive Neuroscience15(4), 559–573 (2003)

    Knauff, M., Fangmeier, T., Ruff, C.C., Johnson-Laird, P.N.: Reasoning, mod- els, and images: behavioral measures and cortical activity. Journal of Cognitive Neuroscience15(4), 559–573 (2003)

    2003

  9. [17]

    Psychological review 18 112, 468–93 (2005)

    Goodwin, G., Johnson-Laird, P.: Reasoning about relations. Psychological review 18 112, 468–93 (2005)

    2005

  10. [18]

    Spatial Cognition & Computation9(2), 109–137 (2009)

    Knauff, M.: A neuro-cognitive theory of deductive relational reasoning with men- tal models and visual images. Spatial Cognition & Computation9(2), 109–137 (2009)

    2009

  11. [19]

    arXiv:2403.15297 (2024) [cs.AI]

    Dong, T., Jamnik, M., Li` o, P.: Sphere Neural-Networks for Rational Reasoning. arXiv:2403.15297 (2024) [cs.AI]

    2024 arXiv

  12. [20]

    In: AAAI (2025)

    Dong, T., Jamnik, M., Li` o, P.: Neural Reasoning for Sure Through Constructing Explainable Models. In: AAAI (2025)

    2025

  13. [21]

    In: Bouamor, H., Pino, J., Bali, K

    Nowak, F., Svete, A., Du, L., Cotterell, R.: On the representational capac- ity of recurrent neural language models. In: Bouamor, H., Pino, J., Bali, K. (eds.) Proceedings of the 2023 Conference on Empirical Methods in Natu- ral Language Processing, pp. 7011–7034. Association ...

    2023 doi

  14. [22]

    Transactions of the Association for Compu- tational Linguistics12, 543–561 (2024) https://doi.org/10.1162/tacl a 00663

    Strobl, L., Merrill, W., Weiss, G., Chiang, D., Angluin, D.: What formal languages can transformers express? a survey. Transactions of the Association for Compu- tational Linguistics12, 543–561 (2024) https://doi.org/10.1162/tacl a 00663

    2024 doi

  15. [23]

    Bloomsbury Academic, ??? (2017)

    Malpass, A., Marfori, M.A.: The History of Philosophical and Formal Logic: From Aristotle to Tarski. Bloomsbury Academic, ??? (2017)

    2017

  16. [24]

    Data and Knowledge Engineering20, 287–303 (1996)

    Smith, B.: Mereotopology: A Theory of Parts and Boundaries. Data and Knowledge Engineering20, 287–303 (1996)

    1996

  17. [25]

    New York, NY:McGraw- Hill, ??? (1981)

    Jeffrey, R.: Formal Logic: Its Scope and Limits (2nd Ed.). New York, NY:McGraw- Hill, ??? (1981)

    1981

  18. [26]

    Psychological Bulletin138(3), 427–457 (2012)

    Khemlani, S., Johnson-Laird, P.N.: Theories of the syllogism: A meta-analysis. Psychological Bulletin138(3), 427–457 (2012)

    2012

  19. [27]

    In: TACAS (1), vol

    Vukmirovic, P., Blanchette, J.C., Cruanes, S., Schulz, S.: Extending a brainiac prover to lambda-free higher-order logic. In: TACAS (1), vol. 11427, pp. 192–210. Springer, ??? (2019)

    2019

  20. [28]

    In: CADE, vol

    Bentkamp, A., Blanchette, J., Tourret, S., Vukmirovic, P.: Superposition for full higher-order logic. In: CADE, vol. 12699, pp. 396–412. Springer, ??? (2021)

    2021

  21. [29]

    Journal of Computer and System Sciences50(1), 132–150 (1995)

    On the computational power of neural nets. Journal of Computer and System Sciences50(1), 132–150 (1995)

    1995

  22. [30]

    In: Diagrams 2018, pp

    Wang, D., Jamnik, M., Li` o, P.: Investigating diagrammatic reasoning with deep neural networks. In: Diagrams 2018, pp. 390–398 (2018) 19

    2018

  23. [31]

    In: ICLR (2020)

    Wang, D., Jamnik, M., Li` o, P.: Abstract diagrammatic reasoning with multiplex graph networks. In: ICLR (2020)

    2020

  24. [32]

    https://arxiv.org/abs/2502.00212

    Dong, K., Ma, T.: STP: Self-play LLM Theorem Provers with Iterative Conjec- turing and Proving (2025). https://arxiv.org/abs/2502.00212

    2025 arXiv

  25. [33]

    https://arxiv.org/abs/2502.07640

    Lin, Y., Tang, S., Lyu, B., Wu, J., Lin, H., Yang, K., Li, J., Xia, M., Chen, D., Arora, S., Jin, C.: Goedel-Prover: A Frontier Model for Open-Source Automated Theorem Proving (2025). https://arxiv.org/abs/2502.07640

    2025 arXiv

  26. [34]

    In: ICLR (2025)

    Li, Z., Liu, H., Zhou, D., Ma, T.: Chain of thought empowers transformers to solve inherently serial problems. In: ICLR (2025)

    2025

  27. [35]

    Google: Palm 2 technical report (2023) arXiv:2305.10403 [cs.CL]

    2023 arXiv

  28. [36]

    https://arxiv.org/abs/2307.09288

    Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bash- lykov, N., Batra, S., Bhargava, P., Bhosale, S., Bikel, D., Blecher, L., Ferrer, C.C., Chen, M., Cucurull, G., Esiobu, D., Fernandes, J., Fu, J., Fu, W., Fuller, B., Gao, C., Goswami, V., Goyal, N...

    2023 arXiv

  29. [37]

    OpenAI: GPT-3.5 (2023)

    2023

  30. [38]

    arXiv:2410.14399 (2025) [cs.CL]

    Wysocka, M., Carvalho, D., Wysocki, O., Valentino, M., Freitas, A.: SylloBio- NLI: Evaluating large language models on biomedical syllogistic reasoning. arXiv:2410.14399 (2025) [cs.CL]

    2025 arXiv

  31. [39]

    Jiang, A.Q., Sablayrolles, A., Mensch, A., Bamford, C., Chaplot, D.S., Casas, D., Bressand, F., Lengyel, G., Lample, G., Saulnier, L., Lavaud, L.R., Lachaux, M.- A., Stock, P., Scao, T.L., Lavril, T., Wang, T., Lacroix, T., Sayed, W.E.: Mistral 7b (2023) arXiv:2310.06825 [cs.CL]

    2023 arXiv

  32. [40]

    Mistral, A.T.: Mixtral of experts- a high quality sparse mixture-of-experts (2023)

    2023

  33. [41]

    https://arxiv.org/abs/2403.08295

    Gemma, T., Google, D.: Gemma: Open Models Based on Gemini Research and Technology (2024). https://arxiv.org/abs/2403.08295

    2024 arXiv

  34. [42]

    MetaAI: The llama 3 herd of models (2024) arXiv:2407.21783 [cs.AI]

    2024 arXiv

  35. [43]

    In: Ku, L.-W., Martins, A., Srikumar, V

    Labrak, Y., Bazoge, A., Morin, E., Gourraud, P.-A., Rouvier, M., Dufour, R.: 20 BioMistral: A collection of open-source pretrained large language models for med- ical domains. In: Ku, L.-W., Martins, A., Srikumar, V. (eds.) Findings of the Association for Computational Linguis...

    2024

  36. [44]

    History and Philosophy of Logic19(1) (1998)

    Hammer, E., Shin, S.J.: Eulers visual logic. History and Philosophy of Logic19(1) (1998)

    1998

  37. [45]

    In: Proceedings - 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022

    He, K., Chen, X., Xie, S., Li, Y., Dollar, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings - 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022. Proceedings of the IEEE Computer Society Conference on Computer Visi...

    2022

  38. [46]

    In: Thirty-seventh Conference on Neural Information Processing Systems (2023)

    Gupta, A., Wu, J., Deng, J., Fei-Fei, L.: Siamese masked autoencoders. In: Thirty-seventh Conference on Neural Information Processing Systems (2023). https://openreview.net/forum?id=yC3q7vInux

    2023

  39. [47]

    MIT Press, Cambridge, MA (2019)

    Simon, H.A.: The Sciences of the Artificial. MIT Press, Cambridge, MA (2019)

    2019

  40. [48]

    Manning, ??? (2024)

    Raschka, S.: Build A Large Language Model (From Scratch). Manning, ??? (2024)

    2024

  41. [49]

    In: Guyon, I., Luxburg, U.V., Bengio, S., Wallach, H., Fergus, R., Vishwanathan, S., Garnett, R

    Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L.u., Polosukhin, I.: Attention is all you need. In: Guyon, I., Luxburg, U.V., Bengio, S., Wallach, H., Fergus, R., Vishwanathan, S., Garnett, R. (eds.) Advances in Neural Information Processi...

    2017

  42. [50]

    Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences478(2022) https://doi.org/10.1098/rspa.2021.0068

    Goyal, A., Bengio, Y.: Inductive biases for deep learning of higher-level cognition. Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences478(2022) https://doi.org/10.1098/rspa.2021.0068

    2022

  43. [51]

    Miller, G.A.: Wordnet: A lexical database for english. Commun. ACM38(11), 39–41 (1995) 21 A The list of 24 valid types of syllogistic reasoning T able 3List of all 24 valid syllogisms, each having a name whose vowels indicate types of moods, e.g., vowels in ‘CE LARENT’ indicat...

    1995

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.