Pith. sign in

REVIEW 3 major objections 5 minor 49 references

Towards a Comparative Framework for Compositional AI Models

T0 review · 3 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 The paper defines compositionality relative to a chosen syntax and uses closure-based tests to show that quantum-circuit and neural DisCoCirc models differ mainly on systematicity.

desk verdict A clean categorical framework for compositionality and a new neural DisCoCirc model, but the headline quantum-vs-neural systematicity gap lacks error bars and rests on a validation scheme that leaks test-distribution data into model selection. read the letter →

arxiv 2507.02940 v1 pith:45OW6EZ5 submitted 2025-06-27 cs.CL cs.AIquant-ph

Tiffany Duneau This is my paper
classification cs.CLcs.AIquant-ph
keywords compositionalgeneralisationDisCoCirccategorytheoryquantumnaturallanguageprocessingneuralnetworksbAbIinterpretabilitysystematicity
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper aims to make compositional generalisation measurable in a way that does not depend on any one model architecture. It defines compositionality relative to a chosen syntax, then adapts four tests — productivity, substitutivity, systematicity, and overgeneralisation — and applies them to two families of models built from the DisCoCirc grammar: quantum circuits and classical neural networks. On a question-answering task derived from bAbI task 6, the two families score within 5% of each other on productivity and substitutivity, but differ by at least 10% on systematicity, where the neural models come closer to random guessing. The paper also shows how a trained compositional model can be interpreted by probing diagram fragments rather than individual boxes, and that this reveals the quantum model effectively answers a simpler question than the one asked. If the framework is adopted, comparisons of compositional generalisation across architectures become direct rather than ad hoc.

What carries the argument

The machinery has three parts. (1) A categorical scaffolding: a language is packaged as a category whose morphisms are stories, a syntax is a functor $e : \Sigma_L \to \mathbf{S}_G$ into a monoidal category of diagrams, and a model is an interpretation functor; a model is syntactically compositional when it factors as $e;g$ and semantically compositional when such a $g$ exists without being specified. The compositional closure $\Sigma^*$ of a training set $\Sigma$ is the set of stories whose diagrams use only the generators appearing in $\Sigma$. (2) A quantitative core: the compositionality factor $\mathrm{cFact}(A,B)=\max(0,\mathrm{acc}_A(M)-\mathrm{acc}_B(M))$ and the compositionality score $\mathrm{cScore}(A,B)=(1-2\,\mathrm{cFact}(A,B))\cdot \mathrm{acc}_A(M)$, which penalise any accuracy lost in moving from a base set $A$ to a harder set $B$. (3) Four closure-based tests — productivity (longer concatenated stories), substitutivity (replacing fragments by synonyms), systematicity (recombining typed fragments), and overgeneralisation (training with corrupted labels) — each defined by a specific choice of $B$ as a closure of $A$. For DisCoCirc, the grammar supplies the syntax, the Sandwich expansion turns higher-order frames into plain boxes, and the two semantic categories are quantum circuits (tensor product) and differentiable functions (direct sum), so the same diagrams can be evaluated in either model.

What would settle it

Re-run the systematicity comparison on a bAbI-6-style task in which object locations matter (the paper notes such a rule can be added) and see whether the quantum-versus-neural gap of at least 10% persists; if the gap disappears, the reported difference is an artifact of the task rather than of the architectures.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central claim is that compositionality is a relational property — a model is compositional relative to a syntax, not absolutely — and that this relational notion can be operationalised as measurable generalisation. A syntax is a functor from stories to a diagram category; a model is syntactically compositional if it factors through that syntax, and semantically compositional if such a factorisation merely exists. The paper turns the four aspects of compositionality into concrete datasets by defining the compositional closure of a training set under each aspect, and measuring the accuracy drop between the base set and its closure. Under this measurement, the quantum and neural DisCoCirc models achieve near-parity on productivity and substitutivity — within 5% — but the systematicity task separates them by at least 10%, with the non-linear neural model overfitting the training data and decaying to chance on the test set. Separately, the paper claims that syntactically compositional models can be interpreted by comparing assertion-relative overlaps of diagram fragments, and it uses this to show that the quantum productivity model has learned to answer 'Is somebody in the park?' rather than 'Is Andrew in the park?', which explains its accuracy loss on longer stories.

Load-bearing premise

The load-bearing premise is that the restricted bAbI 6 fragment and its question-answering task really are compositional under the paper's chosen syntax, so that the train and test splits generated from that syntax have a non-vacuous compositional closure; if the syntax does not match the task's true structure, the reported scores measure something other than compositional generalisation.

Editorial extensions

If this is right

  • Compositionality scores can drive model selection: the Valid AB validation scheme picks models by an estimated compositionality score rather than raw validation accuracy, and it selects models with less overfitting between train and test.
  • Systematicity is the discriminating test among the four aspects; if the result is correct, a benchmark that only measures productivity and substitutivity would miss the main architectural difference.
  • Neural networks can be treated as compositional DisCoCirc models: the direct-sum category of differentiable functions supports the same grammar pipeline and reaches comparable accuracy on two of the three main tasks.
  • Interpretation does not require visualising individual boxes; assertion-relative diagram fragments can explain a model's failure mode, and the paper states this approach extends to models with many qubits per noun and to neural architectures.
  • On overgeneralisation, quantum models follow the compositional rule for corruption levels up to 30% before overfitting, while the neural model overfits uncorrupted samples too, indicating different inductive biases toward rules versus exceptions.

Reading between the lines

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

  • Editorial inference: because the framework relativises compositionality to a syntax, applying it to unrestricted natural language would require a syntax that the paper itself argues is unavailable; in practice the tests will measure restricted fragments, not English as a whole.
  • Editorial inference: the systematicity gap may be rooted in unitarity — reversible quantum circuits preserve distinctions that non-linear feed-forward networks can collapse — but the paper does not test this explanation.
  • Editorial inference: the interpretability result suggests a portable diagnostic for any QA model on this task: check whether it answers the existential question 'Is somebody in X?' instead of the targeted 'Is P in X?'; the paper demonstrates this only for the quantum model.
  • Editorial inference: the same closure-based tests could in principle be applied to transformer-based language models once a syntax for the fragment is fixed, since the categorical definitions do not constrain the model class.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper develops a category-theoretic framework for defining compositional generalisation and compositional interpretability, building on the DisCoCirc architecture. It adapts four tests from Hupkes et al. (productivity, systematicity, substitutivity, overgeneralisation) and introduces the cFact and cScore metrics. These are applied to compare quantum-circuit and classical neural-network DisCoCirc models on a synthetic extension of bAbI task 6. The headline empirical claim is that quantum and neural models score within 5% on productivity and substitutivity but differ by at least 10% on systematicity, with neural models more prone to overfitting. A separate interpretability analysis of a trained quantum productivity model concludes that the model approximates the question 'Is somebody in the park?' rather than the intended per-person question.

Significance. If the empirical comparison were established, the paper would provide a useful comparative benchmark and evidence that tensor-product and direct-sum semantic categories lead to different compositional-generalisation behaviour in DisCoCirc models. The categorical formalisation is a genuine contribution, and the interpretability analysis (Section 5, Appendix G) is a strength: it yields a falsifiable, mechanistic explanation of the productivity failure that is consistent with Figure 12. The paper is also honest about its scope, acknowledging in Section 2.3 that unrestricted English and QA are unlikely to be compositional and that it works in a restricted 'safe corner', and in Section 6 that this is a preliminary investigation. However, the headline quantitative difference currently rests on single selected runs with a validation scheme that leaks test-distribution information, so the main empirical claim is not yet established at the claimed precision.

major comments (3)
  1. [Section 4.3 / Figure 7] The central claim that quantum and neural models differ by at least 10% on systematicity is based on one selected model per architecture, as stated in Section 3.4 ('We train a model each for the quantum, Linear and non-linear neural models'). The only cross-validation analysis is for the productivity task (Section 4.1.1, Figure 8), which shows substantial split-to-split variation for the quantum model. No equivalent variance analysis is reported for systematicity or substitutivity, so the reported 10-point gap cannot be distinguished from run-to-run noise. Please provide multiple seeds or folds with confidence intervals for all four tasks, or explicitly reframe the systematicity comparison as a single-run observation.
  2. [Section 3.4 / Appendix E.1-E.2] The Valid AB validation scheme includes a subset of the test distribution in the validation set and uses a compositionality-score proxy computed on those test-like examples for model selection. This leaks test-distribution information into model choice, which can bias the selected model's absolute scores and, because only one model per architecture is selected, the cross-architecture gap. Appendix E.2 itself reports that the Comp AB and Comp All estimates are 'very volatile' and that selection tends to work only when the true score is above 0.60. This undercuts the reliability of the Figure 7 scores. Please evaluate final models on a genuinely held-out test set that is never used for selection, or use a nested validation procedure that accounts for selection variability.
  3. [Section 2.2, Eqs. (cf)-(cs)] The text says the metric is defined 'for a binary task, with baseline accuracy 0.5', but the baseline does not appear in the equations for cFact or cScore as written. Appendix C.2 reports that the systematicity dataset has a majority-class baseline of 51.6%, not 50%. It should be clarified whether and how the 0.5 baseline enters the score, and whether the reported systematicity scores are adjusted for the actual dataset baseline. This matters because the size of the reported systematicity gap could be sensitive to the baseline convention.
minor comments (5)
  1. [Abstract] The phrases 'within 5%' and 'at least 10%' refer to point estimates from single selected runs; consider adding a qualifier such as 'in this preliminary single-run evaluation' to avoid overprecision.
  2. [Section 2.1 / Figure 3 caption] There are typos: 'Futher examples' should be 'Further examples', and the Figure 3 caption reads 'Notice the the hole'.
  3. [Section 4.1.1 / Figure 8] The cross-validation figure is informative, but the main text should state explicitly that no cross-validation was performed for systematicity or substitutivity; currently this is only inferable from Section 3.4.
  4. [Section 5.2 / Figure 12] The conclusion that the model answers 'Is somebody in the park?' would be strengthened by reporting the full confusion matrix or by showing accuracy separately for 'yes' and 'no' targets with and without confounders; the appendix provides details, but a sentence in the main text would help.
  5. [Appendix C.3 / Appendix D.2 / Table 2] Minor typographical issues: 'repesents' should be 'represents' in Appendix C.3, 'proudctivity' should be 'productivity' in Appendix D.2, and the Table 2 header contains an extra space in 'V ocabulary'.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the metric definitions and empirical comparison are self-contained; the only passage that raises a 'by definition' concern is explicitly flagged and neutralized by an assumption about task labels, not a hidden reduction.

full rationale

The derivation chain in this paper is a sequence of stipulated definitions (cFact, cScore, epsilon-properties, closures) followed by empirical measurements; no parameter is fitted and then renamed as a prediction, and no claimed result is equal to an input by construction. The central behavioural comparison (quantum vs neural systematicity gap) is an observed outcome that could have gone either way. The self-citations to Duneau et al. (2024) supply the quantum ansatz and the seed of the test adaptation, but the formal categorical definitions, the neural implementations, and the experiments are carried out here; no uniqueness theorem or prior result is invoked to forbid alternatives or to force the reported gap. The paper itself flags the one near-circular concern in Section 2.2.3: 'The notion seems to presuppose a syntactically compositional interpretation in order to determine which atoms are synonyms in the first place, which would make the interpretation substitutive by definition.' The paper then neutralizes this by assuming the task labels factor as ∇ = e∇ ; g and defining synonyms from that fixed g, while testing a separate model M against those classes. That is an explicit assumption about the task, not a reduction of the test to itself. Similarly, the Valid AB validation scheme includes test-distribution examples in model selection; this is a disclosed selection-bias/leakage limitation, but it is not circularity because the reported scores are not logically forced by the definitions. Overall, the paper is self-contained against external benchmarks and no circular step is present.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

No free parameters appear in the categorical definitions; the empirical comparison depends on task-specific hyperparameters and training budgets, which are listed above. The core axioms are the compositionality assumptions explicitly flagged in Section 2.3. No new physical or ontological entities are postulated.

free parameters (2)
  • Neural hyperparameters per task = Linear/Linear/Hidden(1); dims 36/12/2; lr 4.97e-3/5.0e-3/50e-3; batch 16/4/1
    Table 5: chosen by Ax tuning or inherited from another task; the comparative claim could depend on these choices and on the tuning budget.
  • Quantum epoch cap = 20-75 epochs, 40-60 in cross-validation
    Appendix D: quantum runs were limited by resource constraints, which may have disadvantaged the quantum models relative to neural models given more training time.
assumptions (3)
  • domain assumption A restricted sub-language of English and the bAbI-style QA task are compositional relative to the chosen syntax e∇
    Section 2.3 states that English and general QA are "unlikely to be compositional" and restricts to a "safe corner". If this restriction fails, the tests do not measure what they claim.
  • domain assumption The bAbI 6 grammar rules are the correct syntax, and any multiple derivations are equivalent
    Section 2.1.1 requires that if there are multiple valid derivations, they are formally equivalent. The dataset construction and all test scores rely on this.
  • domain assumption The overlap between a story circuit and an assertion circuit correctly scores answer candidates
    Section 3.1 defines the model answer as the assertion with greatest overlap; no independent evidence is provided that this is a faithful model of the task.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Towards a Comparative Framework for Compositional AI Models." pith.science (2026). https://pith.science/paper/45OW6EZ5

@misc{pith2026250702940,
  author       = {Pith},
  title        = {Pith review of: Towards a Comparative Framework for Compositional AI Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/45OW6EZ5}},
  note         = {Machine review of arXiv:2507.02940}
}
read the original abstract

The DisCoCirc framework for natural language processing allows the construction of compositional models of text, by combining units for individual words together according to the grammatical structure of the text. The compositional nature of a model can give rise to two things: compositional generalisation -- the ability of a model to generalise outside its training distribution by learning compositional rules underpinning the entire data distribution -- and compositional interpretability -- making sense of how the model works by inspecting its modular components in isolation, as well as the processes through which these components are combined. We present these notions in a framework-agnostic way using the language of category theory, and adapt a series of tests for compositional generalisation to this setting. Applying this to the DisCoCirc framework, we consider how well a selection of models can learn to compositionally generalise. We compare both quantum circuit based models, as well as classical neural networks, on a dataset derived from one of the bAbI tasks, extended to test a series of aspects of compositionality. Both architectures score within 5% of one another on the productivity and substitutivity tasks, but differ by at least 10% for the systematicity task, and exhibit different trends on the overgeneralisation tasks. Overall, we find the neural models are more prone to overfitting the Train data. Additionally, we demonstrate how to interpret a compositional model on one of the trained models. By considering how the model components interact with one another, we explain how the model behaves.

Figures

Figures reproduced from arXiv: 2507.02940 by the authors.

Figure 1
Figure 1. An example syntax derivation for a bAbI 6 story. Each box corresponds to a rule, where the states [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Generator shapes for the syntactic ‘wire’ category of DisCoCirc models. There are three base [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Final state obtained in the DisCoCirc wire category when simplifying the parsing diagram in [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (41 more)
Figure 4
Figure 4. Figure 4: cScore(A, B). Baselines at 0.5 are highlighted with black lines. We define this for a binary task, with baseline accuracy 0.5: cFact(A, B) = max(0, accA(M) − accB(M)) (cf) cScore(A, B) = (1 − 2 · cFact(A, B)) · accA(M) (cs) We visualise this function in [PITH_FULL_IMA…
Figure 5
Figure 5. Figure 5: Ansatze used in this work. Parametrized gates are highlighted in blue, and the circuit diagrams are [PITH_FULL_IMAGE:figures/full_fig_p014_5.png]
Figure 6
Figure 6. Figure 6: Converting an example data entry into a parametrized quantum circuit. Starting from the text [PITH_FULL_IMAGE:figures/full_fig_p015_6.png]
Figure 7
Figure 7. Figure 7: Summary compositionality scores achieved by the best model selected for each task by the quantum [PITH_FULL_IMAGE:figures/full_fig_p017_7.png]
Figure 8
Figure 8. Figure 8: Best selected models according to the Comp AB scheme, and the mean performance across the five splits, for the various model architectures considered. (a) and (b) display the quantum architecture (Sim4 with 3 layers and 1 qubit per noun), where (b) has been trained via…
Figure 9
Figure 9. Figure 9: Visualising the substitutivity models’ accuracy per productivity stratum (context depth). The [PITH_FULL_IMAGE:figures/full_fig_p018_9.png]
Figure 10
Figure 10. Figure 10: Overgeneralisation scores for (a,c) quantum and (b,d) neural models, trained on corrupted datasets. [PITH_FULL_IMAGE:figures/full_fig_p019_10.png]
Figure 11
Figure 11. Figure 11: The assertion-relative fragments evaluated for (ID) the [PITH_FULL_IMAGE:figures/full_fig_p020_11.png]
Figure 12
Figure 12. Figure 12: Model accuracy per number of sentences in the context that include the target location. A [PITH_FULL_IMAGE:figures/full_fig_p021_12.png]
Figure 13
Figure 13. Figure 13: Generic CFG parse, and a finer-grained DisCoCat parse using a pregroup grammar. [PITH_FULL_IMAGE:figures/full_fig_p034_13.png]
Figure 14
Figure 14. Figure 14: Example frames encountered in bAbI 6. Their type signatures are as follows: (a) [PITH_FULL_IMAGE:figures/full_fig_p035_14.png]
Figure 15
Figure 15. Figure 15: Drawing frames with varying amounts of detail. (a) [PITH_FULL_IMAGE:figures/full_fig_p036_15.png]
Figure 16
Figure 16. Figure 16: Depicting the level changing operations as diagrams. The frames displayed have the following [PITH_FULL_IMAGE:figures/full_fig_p036_16.png]
Figure 17
Figure 17. Figure 17: Collapsing frames to the base level. (a) An example frame. (b) Using closure, morphisms can [PITH_FULL_IMAGE:figures/full_fig_p037_17.png]
Figure 18
Figure 18. Figure 18: Generator shapes for the bAbI6 syntax category of DisCoCirc models. [PITH_FULL_IMAGE:figures/full_fig_p038_18.png]
Figure 19
Figure 19. Figure 19: (a) Informal visualisation of the rules as wiring operations for the example derivation given in [PITH_FULL_IMAGE:figures/full_fig_p039_19.png]
Figure 20
Figure 20. Figure 20: (a) The second part of the derivation begun in [PITH_FULL_IMAGE:figures/full_fig_p040_20.png]
Figure 21
Figure 21. Figure 21: An example story and pair of assertions as DisCoCirc diagrams. The binary question [PITH_FULL_IMAGE:figures/full_fig_p041_21.png]
Figure 22
Figure 22. Figure 22: Visualising the balancing for the (a) Productivity and (b) Productivity’ datasets. The stratum in this case is the number of sentences in each story, represented by colour. The distribution of the number of sentences is plotted according to the support depth and answe…
Figure 23
Figure 23. Figure 23: Visualising the distribution of story size, according to the number of sentences and number of [PITH_FULL_IMAGE:figures/full_fig_p042_23.png]
Figure 24
Figure 24. Figure 24: The noun groups used for generating the systematicity dataset. The top and bottom layers [PITH_FULL_IMAGE:figures/full_fig_p043_24.png]
Figure 25
Figure 25. Figure 25: Visualising the splits for the systematicity dataset. The object interactions were not considered [PITH_FULL_IMAGE:figures/full_fig_p043_25.png]
Figure 26
Figure 26. Figure 26: Visualising the distribution of nouns present in each training stratum for the systematicity dataset. [PITH_FULL_IMAGE:figures/full_fig_p044_26.png]
Figure 27
Figure 27. Figure 27: A representation of a structure as a coloured directed acyclic graph. The nouns chosen for the [PITH_FULL_IMAGE:figures/full_fig_p044_27.png]
Figure 28
Figure 28. Figure 28: Visualising the substitutivity dataset. (a) Visualising the balance of splits for the substitutivity [PITH_FULL_IMAGE:figures/full_fig_p045_28.png]
Figure 29
Figure 29. Figure 29: The splits for the Overgeneralisation datasets according to the number of sentences and answer [PITH_FULL_IMAGE:figures/full_fig_p045_29.png]
Figure 30
Figure 30. Figure 30: Visualising the dataset splits for the (a) vanilla and (b) extended validation training regimes. [PITH_FULL_IMAGE:figures/full_fig_p048_30.png]
Figure 31
Figure 31. Figure 31: Productivity scores for the quantum models selected by each validation method, per cross-validation [PITH_FULL_IMAGE:figures/full_fig_p049_31.png]
Figure 32
Figure 32. Figure 32: Comparing the validation and true compositionality scores for the models in each cross-validation [PITH_FULL_IMAGE:figures/full_fig_p051_32.png]
Figure 33
Figure 33. Figure 33: Visualising the accuracy per stratum for the models trained on the (a) productivity, (b) substitu [PITH_FULL_IMAGE:figures/full_fig_p052_33.png]
Figure 34
Figure 34. Figure 34: (a) Computing the overlap between states. Requires a notion of adjoints, or compact closure. The [PITH_FULL_IMAGE:figures/full_fig_p053_34.png]
Figure 35
Figure 35. Figure 35: (a) Simplification of the depiction of moved to, where moved stands for any of the intransitive movement verbs. (b) An assertion-relative diagram fragment. For the bAbI 6 dataset the assertion box can be either is in or is not in. G.1 Box overlaps We visualise the dis…
Figure 36
Figure 36. Figure 36: Visualising the relative overlaps between nouns. On the x-axis, we group the nouns by equivalence [PITH_FULL_IMAGE:figures/full_fig_p054_36.png]
Figure 37
Figure 37. Figure 37: Visualising the relative overlaps between the intransitive [PITH_FULL_IMAGE:figures/full_fig_p054_37.png]
Figure 38
Figure 38. Figure 38: Visualising the assertion-relative overlaps between the transitive [PITH_FULL_IMAGE:figures/full_fig_p055_38.png]
Figure 39
Figure 39. Figure 39: (a) Visualising the magnitude of the overlaps between the diagram fragments (ID), (Ap) (visualised [PITH_FULL_IMAGE:figures/full_fig_p056_39.png]
Figure 40
Figure 40. Figure 40: The selection of single-action diagram fragments considered and their overlaps. (a)-(c) The [PITH_FULL_IMAGE:figures/full_fig_p057_40.png]
Figure 41
Figure 41. Figure 41: The selection of three-noun diagram fragments considered for movement verbs involving a second [PITH_FULL_IMAGE:figures/full_fig_p057_41.png]
Figure 42
Figure 42. Figure 42: Diagram fragments and overlaps for movement combinations involving a second person. (a), (b) [PITH_FULL_IMAGE:figures/full_fig_p058_42.png]
Figure 43
Figure 43. Figure 43: The selection of three noun diagram fragments considered for movement verbs and object [PITH_FULL_IMAGE:figures/full_fig_p058_43.png]
Figure 44
Figure 44. Figure 44: Visualising the magnitude of the assertion-relative overlaps between diagram fragments involving [PITH_FULL_IMAGE:figures/full_fig_p059_44.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

49 extracted references · 21 canonical work pages

  1. [1]

    Alouneh, S

    S. Alouneh, S. Abed, M. H. Al Shayeji, and R. Mesleh. A comprehensive study and analysis on SAT -solvers: advances, usages and achievements. Artificial Intelligence Review, 52 0 (4): 0 2575--2601, Dec. 2019. ISSN 1573-7462. doi:10.1007/s10462-018-9628-0. URL https://doi.org/10.1007/s10462-018-9628-0

    work page doi:10.1007/s10462-018-9628-0 2019

  2. [2]

    Arkoudas

    K. Arkoudas. GPT -4 Can 't Reason . ArXiv, 2023. doi:10.48550/ARXIV.2308.03762. URL https://arxiv.org/abs/2308.03762. Publisher: arXiv Version Number: 2

    arXiv doi:10.48550/arxiv.2308.03762 2023

  3. [3]

    Bakshy, L

    E. Bakshy, L. Dworkin, B. Karrer, K. Kashin, B. Letham, A. Murthy, and S. Singh. Ae: A domain-agnostic platform for adaptive experimentation. In NeurIPS Systems for ML Workshop, 2018. URL http://learningsys.org/nips18/assets/papers/87CameraReadySubmissionAE

    work page 2018

  4. [4]

    Barendregt and H

    H. Barendregt and H. Geuvers. Chapter 18 - Proof - Assistants Using Dependent Type Systems . In A. Robinson and A. Voronkov, editors, Handbook of Automated Reasoning , Handbook of Automated Reasoning , pages 1149--1238. North-Holland, Amsterdam, Jan. 2001. ISBN 978-0-444-50813-3. doi:10.1016/B978-044450813-3/50020-5. URL https://www.sciencedirect.com/scie...

    work page doi:10.1016/b978-044450813-3/50020-5 2001

  5. [5]

    A. R. Barron. Approximation and estimation bounds for artificial neural networks. Machine Learning, 14 0 (1): 0 115--133, Jan. 1994. ISSN 1573-0565. doi:10.1007/BF00993164. URL https://doi.org/10.1007/BF00993164

    doi:10.1007/bf00993164 1994

  6. [6]

    Bubeck, V

    S. Bubeck, V. Chandrasekaran, R. Eldan, J. Gehrke, E. Horvitz, E. Kamar, P. Lee, Y. T. Lee, Y. Li, S. Lundberg, H. Nori, H. Palangi, M. T. Ribeiro, and Y. Zhang. Sparks of Artificial General Intelligence : Early experiments with GPT -4, Apr. 2023. URL http://arxiv.org/abs/2303.12712. arXiv:2303.12712 [cs]

    arXiv 2023

  7. [7]

    N. Chomsky. Aspects of the Theory of Syntax. The MIT Press, 50 edition, 1965. ISBN 9780262527408. URL http://www.jstor.org/stable/j.ctt17kk81z

    work page 1965

  8. [8]

    B. Coecke. The Mathematics of Text Structure . arXiv:1904.03478 [quant-ph], Feb. 2020. URL http://arxiv.org/abs/1904.03478. arXiv: 1904.03478

    arXiv 1904

Show all 49 references
  1. [9]

    Coecke and A

    B. Coecke and A. Kissinger. Picturing Quantum Processes : A First Course in Quantum Theory and Diagrammatic Reasoning . Cambridge University Press, Cambridge, 2017. ISBN 978-1-316-21931-7. doi:10.1017/9781316219317. URL http://ebooks.cambridge.org/ref/id/CBO9781316219317

    2017 doi

  2. [10]

    Coecke, M

    B. Coecke, M. Sadrzadeh, and S. Clark. Mathematical Foundations for a Compositional Distributional Model of Meaning . arXiv:1003.4394 [cs, math], Mar. 2010. URL http://arxiv.org/abs/1003.4394. arXiv: 1003.4394

    2010 arXiv

  3. [11]

    Coecke, G

    B. Coecke, G. de Felice, K. Meichanetzidis, and A. Toumi. Foundations for near-term quantum natural language processing, 2020

    2020

  4. [12]

    Duneau, S

    T. Duneau, S. Bruhn, G. Matos, T. Laakkonen, K. Saiti, A. Pearson, K. Meichanetzidis, and B. Coecke. Scalable and interpretable quantum natural language processing: an implementation on trapped ions, Sept. 2024. URL http://arxiv.org/abs/2409.08777. arXiv:2409.08777

    2024 arXiv

  5. [13]

    Dziri, X

    N. Dziri, X. Lu, M. Sclar, X. L. Li, L. Jiang, B. Y. Lin, P. West, C. Bhagavatula, R. L. Bras, J. D. Hwang, S. Sanyal, S. Welleck, X. Ren, A. Ettinger, Z. Harchaoui, and Y. Choi. Faith and Fate : Limits of Transformers on Compositionality , Oct. 2023. URL http://arxiv.org/abs/...

    2023 arXiv

  6. [14]

    Eldan and O

    R. Eldan and O. Shamir. The Power of Depth for Feedforward Neural Networks , May 2016. URL http://arxiv.org/abs/1512.03965. arXiv:1512.03965

    2016 arXiv

  7. [15]

    Ax - adaptive experimentation platform

    Facebook. Ax - adaptive experimentation platform. https://github.com/facebook/Ax, 2023

    2023

  8. [16]

    M. M. Ferdaus, M. Abdelguerfi, E. Ioup, K. N. Niles, K. Pathak, and S. Sloan. Towards Trustworthy AI : A Review of Ethical and Robust Large Language Models . ArXiv, 2024. doi:10.48550/ARXIV.2407.13934. URL https://arxiv.org/abs/2407.13934. Publisher: arXiv Version Number: 1

    2024 doi arXiv

  9. [17]

    J. A. Fodor and Z. W. Pylyshyn. Connectionism and cognitive architecture: A critical analysis. Cognition, 28 0 (1): 0 3--71, Mar. 1988. ISSN 0010-0277. doi:10.1016/0010-0277(88)90031-5. URL https://www.sciencedirect.com/science/article/pii/0010027788900315

    1988

  10. [18]

    Glasser, R

    I. Glasser, R. Sweke, N. Pancotti, J. Eisert, and J. Cirac. Expressive power of tensor-network factorizations for probabilistic modeling, with applications from hidden Markov models to quantum machine learning. In Neural Information Processing Systems, July 2019. URL https://w...

    2019

  11. [19]

    R. F. Hadley. Systematicity in Connectionist Language Learning . Mind & Language, 9 0 (3): 0 247--272, 1994. ISSN 1468-0017. doi:10.1111/j.1468-0017.1994.tb00225.x. URL https://onlinelibrary.wiley.com/doi/abs/10.1111/j.1468-0017.1994.tb00225.x. \_eprint: https://onlinelibrary....

    1994

  12. [20]

    Hupkes, V

    D. Hupkes, V. Dankers, M. Mul, and E. Bruni. Compositionality decomposed: how do neural networks generalise?, Feb. 2020. URL http://arxiv.org/abs/1908.08351. arXiv:1908.08351 [cs, stat]

    2020 arXiv

  13. [21]

    Johnson and Journal of Philosophy, Inc

    K. Johnson and Journal of Philosophy, Inc. On the Systematicity of Language and Thought :. Journal of Philosophy, 101 0 (3): 0 111--139, 2004. ISSN 0022-362X. doi:10.5840/jphil2004101321. URL http://www.pdcnet.org/oom/service?url_ver=Z39.88-2004&rft_val_fmt=&rft.imuse_id=jphil...

    2004 doi

  14. [22]

    Kartsaklis, I

    D. Kartsaklis, I. Fan, R. Yeung, A. Pearson, R. Lorenz, A. Toumi, G. de Felice, K. Meichanetzidis, S. Clark, and B. Coecke. lambeq: An efficient high-level python library for quantum N L P . arXiv:2110.04236, 2021

    2021 arXiv

  15. [23]

    Karttunen

    L. Karttunen. Discourse Referents . In International Conference on Computational Linguistics COLING 1969: Preprint No . 70 , Sånga Säby, Sweden, Sept. 1969. URL https://aclanthology.org/C69-7001

    1969

  16. [24]

    D. P. Kingma and J. Ba. Adam: A method for stochastic optimization. In Y. Bengio and Y. LeCun, editors, 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings , 2015. URL http://arxiv.org/abs/1412.6980

    2015 arXiv

  17. [25]

    Krawchuk, N

    C. Krawchuk, N. Khatri, N. J. Ortega, and D. Kartsaklis. Efficient Generation of Parameterised Quantum Circuits from Large Texts , May 2025. URL http://arxiv.org/abs/2505.13208. arXiv:2505.13208 [quant-ph]

    2025 arXiv

  18. [26]

    Laakkonen, K

    T. Laakkonen, K. Meichanetzidis, and B. Coecke. Quantum Algorithms for Compositional Text Processing . Electronic Proceedings in Theoretical Computer Science, 406: 0 162--196, Aug. 2024. ISSN 2075-2180. doi:10.4204/EPTCS.406.8. URL http://arxiv.org/abs/2408.06061. arXiv:2408.0...

    2024 arXiv

  19. [27]

    J. Lambek. Pregroups and natural language processing. The Mathematical Intelligencer, 28 0 (2): 0 41--48, Mar. 2006. ISSN 0343-6993. doi:10.1007/BF02987155. URL https://doi.org/10.1007/BF02987155

    2006 doi

  20. [28]

    Leinster

    T. Leinster. Higher Operads , Higher Categories , May 2003. URL http://arxiv.org/abs/math/0305049. arXiv:math/0305049

    2003 arXiv

  21. [29]

    J. Liu, R. A. Shaikh, B. Rodatz, R. Yeung, and B. Coecke. A Pipeline For Discourse Circuits From CCG , Nov. 2023. URL http://arxiv.org/abs/2311.17892. arXiv:2311.17892 [cs]

    2023 arXiv

  22. [30]

    Lorenz, A

    R. Lorenz, A. Pearson, K. Meichanetzidis, D. Kartsaklis, and B. Coecke. Qnlp in practice: Running compositional models of meaning on a quantum computer. Journal of Artificial Intelligence Research, 76: 0 1305--1342, 04 2023. doi:10.1613/jair.1.14329

    2023 doi

  23. [31]

    Mirzadeh, K

    I. Mirzadeh, K. Alizadeh, H. Shahrokhi, O. Tuzel, S. Bengio, and M. Farajtabar. GSM - Symbolic : Understanding the Limitations of Mathematical Reasoning in Large Language Models , Oct. 2024. URL http://arxiv.org/abs/2410.05229. arXiv:2410.05229 [cs]

    2024 arXiv

  24. [32]

    Achiam, S

    OpenAI, J. Achiam, S. Adler, S. Agarwal, L. Ahmad, I. Akkaya, F. L. Aleman, D. Almeida, J. Altenschmidt, S. Altman, S. Anadkat, R. Avila, I. Babuschkin, S. Balaji, V. Balcom, P. Baltescu, H. Bao, M. Bavarian, J. Belgum, I. Bello, J. Berdine, G. Bernadett-Shapiro, C. Berner, L....

    2024 arXiv

  25. [33]

    B. Partee. Compositionality. Varieties of formal semantics, Jan. 1984. URL https://www.academia.edu/3187795/Compositionality

    1984

  26. [34]

    Paszke, S

    A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga, A. Desmaison, A. Kopf, E. Yang, Z. DeVito, M. Raison, A. Tejani, S. Chilamkurthy, B. Steiner, L. Fang, J. Bai, and S. Chintala. PyTorch: An Imperative Style, High-Per...

    2019

  27. [35]

    Pelletier

    F. Pelletier. The Principle of Semantic Compositionality . Topoi, 13: 0 11--24, Mar. 1994. doi:10.1007/BF00763644

    1994 doi

  28. [36]

    M. Pentus. Lambek grammars are context free. In [1993] Proceedings Eighth Annual IEEE Symposium on Logic in Computer Science , pages 429--433, June 1993. doi:10.1109/LICS.1993.287565

    1993

  29. [37]

    Roberts, A

    C. Roberts, A. Milsted, M. Ganahl, A. Zalcman, B. Fontaine, Y. Zou, J. Hidary, G. Vidal, and S. Leichenauer. Tensornetwork: A library for physics and machine learning. arXiv:1905.01330, 2019

    1905 arXiv

  30. [38]

    C. Rudin. Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nature Machine Intelligence, 1 0 (5): 0 206--215, May 2019. ISSN 2522-5839. doi:10.1038/s42256-019-0048-x. URL https://www.nature.com/articles/s42256-019...

    2019 doi

  31. [39]

    Selinger

    P. Selinger. A survey of graphical languages for monoidal categories. In B. Coecke, editor, New Structures for Physics, Lecture Notes in Physics, pages 275--337. Springer-Verlag, 2011. a rXiv:0908.3347

    2011 arXiv

  32. [40]

    S. Sim, P. D. Johnson, and A. Aspuru‐Guzik. Expressibility and entangling capability of parameterized quantum circuits for hybrid quantum‐classical algorithms. Advanced Quantum Technologies, 2 0 (12), Oct. 2019. ISSN 2511-9044. doi:10.1002/qute.201900070. URL http://dx.doi.org...

    2019 doi

  33. [41]

    Surace and L

    J. Surace and L. Tagliacozzo. Fermionic Gaussian states: an introduction to numerical approaches. SciPost Physics Lecture Notes, page 54, May 2022. ISSN 2590-1990. doi:10.21468/SciPostPhysLectNotes.54. URL http://arxiv.org/abs/2111.08343. arXiv:2111.08343 [quant-ph]

    2022 arXiv

  34. [42]

    Z. G. Szabó. The case for compositionality. In W. Hinzen, E. Machery, and M. Werning, editors, The Oxford Handbook of Compositionality , page 0. Oxford University Press, Feb. 2012. ISBN 978-0-19-954107-2. doi:10.1093/oxfordhb/9780199541072.013.0003. URL https://doi.org/10.1093...

    2012

  35. [43]

    Tamari, K

    R. Tamari, K. Richardson, A. Sar-Shalom, N. Kahlon, N. Liu, R. Tsarfaty, and D. Shahaf. Dyna- bAbI : unlocking bAbI 's potential with dynamic synthetic benchmarking, Nov. 2021. URL http://arxiv.org/abs/2112.00086. arXiv:2112.00086

    2021 arXiv

  36. [44]

    Touvron, T

    H. Touvron, T. Lavril, G. Izacard, X. Martinet, M.-A. Lachaux, T. Lacroix, B. Rozière, N. Goyal, E. Hambro, F. Azhar, A. Rodriguez, A. Joulin, E. Grave, and G. Lample. LLaMA : Open and Efficient Foundation Language Models , Feb. 2023. URL http://arxiv.org/abs/2302.13971. arXiv...

    2023 arXiv

  37. [45]

    S. Tull, R. Lorenz, S. Clark, I. Khan, and B. Coecke. Towards Compositional Interpretability for XAI , June 2024. URL http://arxiv.org/abs/2406.17583. arXiv:2406.17583 [cs, math]

    2024 arXiv

  38. [46]

    Vilone and L

    G. Vilone and L. Longo. Notions of explainability and evaluation approaches for explainable artificial intelligence. Information Fusion, 76: 0 89--106, Dec. 2021. ISSN 1566-2535. doi:10.1016/j.inffus.2021.05.009. URL https://www.sciencedirect.com/science/article/pii/S1566253521001093

    2021 doi

  39. [47]

    Weston, A

    J. Weston, A. Bordes, S. Chopra, A. M. Rush, B. van Merriënboer, A. Joulin, and T. Mikolov. Towards A I -complete question answering: A set of prerequisite toy tasks. arXiv:1502.05698, 2015

    2015 arXiv

  40. [48]

    Yahya, K

    M. Yahya, K. Berberich, S. Elbassuoni, M. Ramanath, V. Tresp, and G. Weikum. Natural Language Questions for the Web of Data . In J. Tsujii, J. Henderson, and M. Paşca, editors, Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language Processing and Com...

    2012

  41. [49]

    Zeng and B

    W. Zeng and B. Coecke. Quantum algorithms for compositional natural language processing. Electronic Proceedings in Theoretical Computer Science, 221: 0 67–75, Aug. 2016. ISSN 2075-2180. doi:10.4204/eptcs.221.8. URL http://dx.doi.org/10.4204/EPTCS.221.8

    2016 doi

Pith tools

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