Pith. sign in

REVIEW 3 major objections 5 minor 21 references

Compositional Concept Generalization with Variational Quantum Circuits

T0 review · 3 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read Quantum-trained DisCoCat representations beat classically trained ones on out-of-distribution image captions.

desk verdict A useful proof-of-concept, but the central quantum-vs-classical comparison is confounded: the trainable image box, not quantum training, likely explains the reported advantage. read the letter →

arxiv 2509.09541 v1 pith:LJUC2ELR submitted 2025-09-11 cs.AI

Hala Hawashin , Mina Abbaszadeh , Nicholas Joseph , Beth Pearson , Martha Lewis , Mehrnoosh sadrzadeh This is my paper
classification cs.AI
keywords compositionalgeneralizationvariationalquantumcircuitsDisCoCatnaturallanguageprocessingout-of-distributionimagecaptioningcontrastivelearningrelationalreasoning
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 asks whether representing compositional sentence meanings as quantum circuits can improve compositional generalization—the ability to handle unseen combinations of known words. It implements the DisCoCat framework, where grammar dictates how word vectors and tensors contract, as variational quantum circuits, and trains them on an image-captioning task where the model must decide which spatial caption matches a picture. On the out-of-distribution test split, quantum-trained representations beat classically trained DisCoCat representations by a wide margin (64.06% vs 30.63% with multi-hot image encodings; 50.31% vs 0% with CLIP vectors), while using hundreds of trainable parameters compared with CLIP's tens of millions. The authors conclude that quantum models are less prone to overfitting when learning compositional tensors and offer a proof of concept for quantum natural language processing on multimodal tasks.

What carries the argument

The central machinery is a DisCoCat-to-quantum translation: compact-closed categorical string diagrams for sentences are turned into IQP variational circuits, with grammatical 'cups' implemented as postselected contractions, nouns as qubit registers, and relation words as parameterized rank-2 and rank-3 tensors. Images are loaded as quantum states either by multi-hot binary encoding or by angle/amplitude encoding of PCA-reduced CLIP vectors. To compare image and sentence circuits, the paper either appends a trainable 'unifying box' that reduces the image to one qubit or pads the sentence space to the image dimension. The matching score is the inner product between the two circuit outputs, tr

What would settle it

Train a classical DisCoCat model on the exact same simplified captions and multi-hot image encodings with the same 100 epochs and a modest learning-rate grid search; if it matches or exceeds Quantum-MHE's 64.06% OOD test accuracy, the claim that quantum training efficiency causes the reduced overfitting is unsupported.

Watch

Extended reading notes

Core claim

The central claim is that translating DisCoCat's tensor-based compositional semantics into variational quantum circuits yields representations that generalize out-of-distribution better than the same semantics trained classically. Grammatical structure is wired directly into the circuit; words and relation words are parameterized rotations; images enter as quantum states, either as binary multi-hot vectors or as PCA-reduced CLIP embeddings; and the inner product between image and sentence circuit outputs is the matching score optimized by a contrastive loss. The strongest evidence is Quantum-MHE reaching 64.06% on the OOD test split versus 30.63% for Classic-DisCoCat-MHE, and Quantum-CLIP wi

Load-bearing premise

The load-bearing premise is that the simplified caption form 'noun isLeftOf noun' preserves all the semantic structure needed to test compositional generalization; if the dropped grammatical material matters, the benchmark no longer tests the phenomenon the paper claims to measure.

Editorial extensions

If this is right

  • Quantum-trained DisCoCat representations generalize to unseen combinations of known shapes and relations substantially better than classically trained DisCoCat, so the viability of compositional tensor semantics should be re-evaluated rather than dismissed from earlier negative results.
  • Because quantum models achieve this with a few hundred parameters and 100 epochs, training efficiency, not model capacity, is the most plausible cause of the improved out-of-distribution performance.
  • The choice of image encoding is decisive: multi-hot encodings give strong results, while CLIP vectors expose shape-level failures (sphere is never recognized), indicating that future gains require staged training of shape features before relations.
  • On the unseen validation split, Quantum-CLIP with angle encoding (61.5%) approaches fine-tuned CLIP (63%), suggesting that near-parity with much larger models may be achievable on this task with better training procedures rather than more parameters.

Reading between the lines

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

  • Editorial: if the training-efficiency explanation is right, the quantum advantage should grow with tensor rank or sentence length, because classical tensor contraction costs grow exponentially while circuit depth grows linearly in qubits; a controlled scaling experiment over relations with more arguments would test this.
  • Editorial: the difference between angle encoding (50.31%) and amplitude encoding (41.25%) suggests that the inductive bias of the classical-to-quantum embedding matters more than raw circuit expressiveness; amplitude encoding entangles all features immediately, which may make the later parameterized layers work harder to single out the subject–relation–object structure.
  • Editorial: the simplification to 'noun isLeftOf noun' keeps the benchmark tractable but removes determiners and verb phrase structure; if the DisCoCat-to-circuit translation is ever claimed to handle natural captions, the 27k-gates-per-caption complexity of the full sentence form must be addressed by circuit compression or factorization, not by simplification.
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. This manuscript presents an empirical study of whether Variational Quantum Circuits (VQCs) improve compositional generalization in a tensor-based DisCoCat model. The authors simplify captions to 'noun isLeftOf/RightOf noun', encode images with either multi-hot encodings or CLIP vectors (angle/amplitude encoding), and train IQP-style VQCs to match image and sentence output states. They compare against classically trained DisCoCat models and a CLIP baseline. The central reported finding is that quantum-trained DisCoCat models outperform classical DisCoCat models on the OOD test split (e.g., Quantum-MHE 64.06% vs Classic-DisCoCat-MHE 30.63%; Quantum-CLIP angle 50.31% vs Classic-DisCoCat-CLIP 0.00%), which the authors interpret as evidence that quantum models are less susceptible to overfitting on the training set.

Significance. If the advantage were cleanly attributable to the quantum parameterization, this would be an interesting proof of concept with very small trainable parameter counts. The paper is clearly organized, connects categorical semantics to a concrete quantum circuit implementation, and uses a task simple enough to be reproducible. It also honestly frames the work around a conjecture rather than a theorem. However, the central attribution is undermined by the trainable image box used only in the quantum models, by the seed-selection and reporting choices, and by asymmetric training budgets. The paper's own Table IV suggests the quantum advantage may largely disappear without the trainable image box. With additional controlled experiments, the comparison could become informative for the quantum-NLP community.

major comments (3)
  1. [Section III-B, Section III-C, Table III] The headline comparison is confounded. The 'Unifying Trainable Image Box' used for all quantum models in Table III has trainable rotation gates, so 'we are in effect training the image vectors' (Section III-B). The classical DisCoCat baselines in Section III-C instead use fixed MHE or CLIP image vectors and train only the relation tensors. Thus Quantum-MHE 64.06% vs Classic-DisCoCat-MHE 30.63% and Quantum-CLIP-angle 50.31% vs Classic-DisCoCat-CLIP 0.00% change two variables at once. Table IV, where image vectors are not trained, shows near-chance quantum performance (50.00% and 48.75%), suggesting that most of the advantage in Table III may be due to the trainable image-side projection. A classical DisCoCat trained with a small trainable image projection of comparable capacity is needed to attribute the improvement to the quantum training rather than to this extra trainable component.
  2. [Section IV, Tables III and IV] The paper reports only the seed with the best OOD validation accuracy from four random seeds and gives no error bars or variance. With a chance level of 50%, Quantum-CLIP angle 50.31% and Quantum-MHE 50.00% in Table IV are statistically indistinguishable from chance. The claim that quantum models 'consistently outperform' classical DisCoCat is not supported without reporting all seeds, means/standard deviations, and ideally significance tests. The 0.00% test accuracy for Classic-DisCoCat-CLIP (with 66.00% validation) also needs explanation; if this is a single unlucky seed, the comparison is especially fragile.
  3. [Section I, Section IV] Quantum and classical models are trained with asymmetric budgets: 100 epochs for quantum vs 50 for classical (Section IV). Learning rates also differ (0.001 for quantum, 0.1/0.01 for classical DisCoCat). The paper's motivating conjecture is increased training efficiency of quantum models, but the experiments do not compare efficiency under matched budgets. To substantiate 'less susceptible to overfitting' or 'train better', the authors should add matched-epoch comparisons or accuracy-versus-epoch curves. As reported, the results cannot distinguish an efficiency/inductive-bias effect from simply allowing the quantum models twice as many training updates.
minor comments (5)
  1. [Section IV] Typo: 'Acorss all models' should be 'Across all models'.
  2. [Section III-B, Section IV] Please reconcile '27k gates per caption' with the later statement that 'each caption is encoded by a variational circuit with 1 qubit per NOUN and 3 layers, yielding 36 trainable parameters per caption'. These statements are hard to align with the 8-bit MHE vectors and the 9/12-qubit CLIP encodings.
  3. [Tables III and IV] The 'noise' injected into MHE encodings is never defined in Sections III-B or IV. Please specify what noise is added and why it is used.
  4. [Section III-B] The claim that the simplified form 'noun is{left/right}Of noun' retains 'all the essential semantic information of a caption' is too strong. The simplification may be useful for tractability, but it drops grammatical structure and should be presented as a limitation rather than as an assertion of semantic equivalence.
  5. [Tables III and IV] The CLIP Frozen/Fine-tuned rows are not DisCoCat baselines and have vastly larger parameter counts. Clearly separate these reference points from the DisCoCat comparisons in the text and figure captions.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper is an empirical comparison whose central claims are tested on held-out OOD data, not derived from the inputs.

full rationale

The paper's central claim is an empirical conjecture—that quantum-trained DisCoCat representations outperform classically trained ones on an OOD image-captioning split. This is tested by training multiple models and reporting held-out test accuracies (Tables III and IV), not by deriving the result from the definition of the models or from fitted parameters. The classical baselines are implemented from [7], which includes a co-author, and the dataset comes from [14], also co-authored, but these are used as experimental scaffolds and baselines, not as premises that logically force the quantum advantage. The 'Unifying Trainable Image Box' does train the image vectors ('we are in effect training the image vectors'), which differs from the classical baselines' fixed image vectors; this is a potential confound or validity threat for the attribution of the advantage, but it is not circular reasoning—the test accuracy is still an independent empirical outcome, not a quantity defined in terms of the target claim. The simplification to 'noun isLeftOf noun' is an assumption about task sufficiency, not a self-definitional loop. The IQP ansatz is adopted from prior work as an ansatz, not presented as a derived first-principles result. No fitted parameter is renamed as a prediction, and no uniqueness theorem is imported from the authors' own work to force a choice. Therefore the paper exhibits no circular step under the definitions used here; the confounding comparison is a correctness/experimental-design issue, not circularity.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The paper contributes no derivation; everything rests on the DisCoCat-to-VQC mapping from prior work, a hand-chosen sentence simplification, and standard contrastive training. The main free parameters beyond standard training are the number of qubits/PCA dimension and the unexplained MHE noise.

free parameters (3)
  • MHE noise injection
    Best results use 'noisy' MHE (64.06% test) but the noise model and amplitude are never specified, so the headline number is not reproducible.
  • CLIP PCA dimension / number of qubits = 3, 9, 12 qubits
    Chosen by hand for angle and amplitude encodings (Section III-B); directly controls information loaded into the circuit.
  • Training hyperparameters = quantum LR 0.001, 100 epochs; classical LR 0.1/0.01/1e-5, 50 epochs
    Quantum and classical models are trained with different budgets and selected via hyperparameter search, making the comparison not apples-to-apples.
assumptions (5)
  • domain assumption Compact closed categories with pregroup grammars model grammatical composition (Section II-A, after [1], [16]).
    The entire meaning construction assumes that pregroup reductions correspond to tensor contractions in FHilb.
  • domain assumption FHilb is an appropriate meaning category and inner products measure semantic similarity (Section II-A).
    Sentence-image matching scores are inner products; this identifies semantic similarity with Hilbert-space geometry.
  • domain assumption The IQP ansatz with post-selected cups implements DisCoCat diagram contraction (Section II-B, from [10]).
    VQCs are taken to faithfully realize the tensor operations; no noise-model or hardware validation is given.
  • ad hoc to paper The simplified caption form 'noun isLeftOf noun' retains all semantic information needed for the task (Section III-B).
    The full sentences are reduced to this form to lower circuit depth; if the simplification removes compositional structure, the benchmark does not test the claimed phenomenon.
  • domain assumption Contrastive loss on correct/incorrect caption pairs yields a valid training signal (Section III-C).
    Standard contrastive learning assumption: increasing similarity of matching image-caption pairs and decreasing nonmatching pairs learns representations.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Compositional Concept Generalization with Variational Quantum Circuits." pith.science (2026). https://pith.science/paper/LJUC2ELR

@misc{pith2026250909541,
  author       = {Pith},
  title        = {Pith review of: Compositional Concept Generalization with Variational Quantum Circuits},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LJUC2ELR}},
  note         = {Machine review of arXiv:2509.09541}
}
read the original abstract

Compositional generalization is a key facet of human cognition, but lacking in current AI tools such as vision-language models. Previous work examined whether a compositional tensor-based sentence semantics can overcome the challenge, but led to negative results. We conjecture that the increased training efficiency of quantum models will improve performance in these tasks. We interpret the representations of compositional tensor-based models in Hilbert spaces and train Variational Quantum Circuits to learn these representations on an image captioning task requiring compositional generalization. We used two image encoding techniques: a multi-hot encoding (MHE) on binary image vectors and an angle/amplitude encoding on image vectors taken from the vision-language model CLIP. We achieve good proof-of-concept results using noisy MHE encodings. Performance on CLIP image vectors was more mixed, but still outperformed classical compositional models.

Figures

Figures reproduced from arXiv: 2509.09541 by the authors.

Figure 1
Figure 1. Sentence derivation in the graphical calculus. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 3
Figure 3. A variational quantum circuit for the sentence in [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 5
Figure 5. Dataset design. Class labels belonging to each dataset [PITH_FULL_IMAGE:figures/full_fig_p004_5.png] view at source ↗
Figures from the paper (4 more)
Figure 8
Figure 8. Figure 8: The VQC of ‘Sphere isLeftOf cylinder’. one is represented by 4 bits , resulting in an 8 bits vectors for the whole caption. The resulting embeddings are shown in tables I and II. MHE’s focus on the essential data required to solve a task and discard other details typic…
Figure 7
Figure 7. Figure 7: The diagrammatic representation of ‘Sphere isLeftOf [PITH_FULL_IMAGE:figures/full_fig_p005_7.png]
Figure 9
Figure 9. Figure 9: The trainable box. C. Classical experiments We also implement the classical DisCoCat representations of relations and sentences using techniques from [7]. In these representations the relations isLeftOf and isRightOf are represented as matrices in R d×d , where d is th…
Figure 10
Figure 10. Figure 10: Diagram of the subject, relation, and object represen￾tations for the classical representations. Subject, object and relation representations are implemented in PyTorch and trained using contrastive learning. Specifically, we minimize the following quantity using grad…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

21 extracted references · 1 canonical work pages

  1. [1]

    Mathematical foundations for a compositional distributional model of meaning,

    B. Coecke, M. Sadrzadeh, and S. Clark, “Mathematical foundations for a compositional distributional model of meaning,”arXiv:1003.4394 [cs, math], Mar. 2010. [Online]. Available: http://arxiv.org/abs/1003.4394

    arXiv 2010

  2. [2]

    A vector space model for automatic indexing,

    G. Salton, A. Wong, and C. S. Yang, “A vector space model for automatic indexing,”Communications of the ACM, vol. 18, no. 11, pp. 613–620, 1975

    1975

  3. [3]

    A solution to Plato’s problem: the latent semantic analysis theory of acquisition, induction, and representa- tion of knowledge,

    T. K. Landauer and S. T. Dumais, “A solution to Plato’s problem: the latent semantic analysis theory of acquisition, induction, and representa- tion of knowledge,”Psychological Review, vol. 104, no. 2, pp. 211–240, 1997

    1997

  4. [4]

    Distributed representations of words and phrases and their composition- ality,

    T. Mikolov, I. Sutskever, K. Chen, G. S. Corrado, and J. Dean, “Distributed representations of words and phrases and their composition- ality,” inAdvances in Neural Information Processing Systems (NeurIPS), vol. 26. Curran Associates, Inc., 2013, pp. 3111–3119

    2013

  5. [5]

    Learning adjective meanings with a tensor-based skip-gram model,

    J. Maillard and S. Clark, “Learning adjective meanings with a tensor-based skip-gram model,” inProceedings of the Nineteenth Conference on Computational Natural Language Learning. Beijing, China: Association for Computational Linguistics, Jul. 2015, pp. 327–331. [Online]. Available: https://aclanthology.org/K15-1035/

    2015

  6. [6]

    Representation learning for type-driven composition,

    G. Wijnholds, M. Sadrzadeh, and S. Clark, “Representation learning for type-driven composition,” inProceedings of the 24th Conference on Computational Natural Language Learning (CoNLL). Online: Association for Computational Linguistics, Nov. 2020, pp. 313–324. [Online]. Available: https://aclanthology.org/2020.conll-1.24

    2020

  7. [7]

    Does CLIP bind concepts? probing compositionality in large image models,

    M. Lewis, N. Nayak, P. Yu, J. Merullo, Q. Yu, S. Bach, and E. Pavlick, “Does CLIP bind concepts? probing compositionality in large image models,” inFindings of the Association for Computational Linguistics: EACL 2024, Y . Graham and M. Purver, Eds. St. Julian’s, Malta: Association for Computational Linguistics, Mar. 2024, pp. 1487–1500

    2024

  8. [8]

    A categorical semantics of quantum protocols,

    S. Abramsky and B. Coecke, “A categorical semantics of quantum protocols,” 2007. [Online]. Available: https://arxiv.org/abs/quant- ph/0402130

    arXiv 2007

Show all 21 references
  1. [9]

    Lambeq: An efficient high-level python library for quantum NLP,

    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 NLP,”arXiv:2110.04236 [quant-ph], Oct. 2021. [Online]. Available: http://arxiv.org/abs/2110.04236

    2021 arXiv

  2. [10]

    QNLP in practice: running compositional models of meaning on a quantum computer,

    R. Lorenz, A. Pearson, K. Meichanetzidis, D. Kartsaklis, and B. Coecke, “QNLP in practice: running compositional models of meaning on a quantum computer,”arXiv:2102.12846 [quant-ph], Feb. 2021. [Online]. Available: http://arxiv.org/abs/2102.12846

    2021 arXiv

  3. [11]

    Foundations for near-term quantum natural language processing,

    B. Coecke, G. de Felice, K. Meichanetzidis, and A. Toumi, “Foundations for near-term quantum natural language processing,” 2020. [Online]. Available: https://arxiv.org/abs/2012.03755

    2020 arXiv

  4. [12]

    Grammar- aware sentence classification on quantum computers,

    K. Meichanetzidis, A. Toumi, G. de Felice, and B. Coecke, “Grammar- aware sentence classification on quantum computers,”Quantum Machine Intelligence, vol. 5, no. 1, Feb. 2023

    2023

  5. [13]

    Large scale structure-aware pronoun resolution using quantum natural language processing,

    H. Wazni, K. I. Lo, L. McPheat, and M. Sadrzadeh, “Large scale structure-aware pronoun resolution using quantum natural language processing,”Quantum Machine Intelligence, vol. 6, p. 60, 2024. [Online]. Available: https://doi.org/10.1007/s42484-024-00193-w

    2024 doi

  6. [14]

    Diffusion models for improved compositional generalisation in VLMs,

    B. Pearson, M. Wray, and M. Lewis, “Diffusion models for improved compositional generalisation in VLMs,” Proc. 2nd Workshop on What is Next in Multimodal Foundation Models?, 2024

    2024

  7. [15]

    Learning transferable visual models from natural language supervi- sion,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, G. Krueger, and I. Sutskever, “Learning transferable visual models from natural language supervi- sion,” inProceedings of the 38th International Conference on Machine ...

  8. [16]

    Bell states and negative sentences in the distributed model of meaning,

    A. Preller and M. Sadrzadeh, “Bell states and negative sentences in the distributed model of meaning,”Electronic Notes in Theoretical Computer Science, vol. 270, no. 2, pp. 141–153, 2011

    2011

  9. [17]

    A survey of graphical languages for monoidal categories,

    P. Selinger, “A survey of graphical languages for monoidal categories,” inNew Structures for Physics. Springer, 2010, pp. 289–355. [Online]. Available: https://arxiv.org/abs/0908.3347

    2010 arXiv

  10. [18]

    M. A. Nielsen and I. L. Chuang,Quantum Computation and Quantum Information. Cambridge, UK: Cambridge University Press, 2010

    2010

  11. [19]

    Classification of the MNIST data set with quantum slow feature analysis,

    I. Kerenidis and A. Luongo, “Classification of the MNIST data set with quantum slow feature analysis,”Phys. Rev. A, vol. 101, no. 6, p. 062327, Jun. 2020

    2020

  12. [20]

    Reasoning about meaning in natural language with compact closed categories and Frobenius algebras,

    D. Kartsaklis, M. Sadrzadeh, S. Pulman, and B. Coecke, “Reasoning about meaning in natural language with compact closed categories and Frobenius algebras,” inLogic and Algebraic Structures in Quantum Computing, 2013, p. 199

    2013

  13. [139]

    2021, pp

    PMLR, Jul. 2021, pp. 8748–8763

    2021

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.