Official implementation of "Meaning Representations as Variational Quantum Circuits"
Tilen G. Limbäck-Stokin, Tanishka A. Birdavade, Kin Ian Lo, Mehrnoosh Sadrzadeh > Quantum Learning Labs, University College London (UCL)
Classical Vision-Language Models (VLMs) like CLIP rely heavily on unstructured sequences and
By treating grammatical compositions as quantum entangling operations and words as parameterized unitary rotations, QuLIP achieves competitive multimodal alignment (e.g., 83.16% on SVO-Swap) while utilizing two orders of magnitude fewer parameters (10k-100k) than classical baselines like OpenCLIP (63M).
Figure 1: The QuLIP multimodal pipeline. Images are classically embedded and amplitude-encoded. Sentences are parsed via CCG and topologically mapped to VQCs to generate the text state. Alignment is computed via a quantum inner product.
The codebase is modularized to support CCG parsing, topological circuit compilation, scalable tensor contraction, and custom quantum loss landscapes.
data_processing.py: Handles multimodal dataset ingestion (ARO, SVO-Swap), CLIP image extraction, and maps classical embeddings to PennyLaneAmplitudeEmbeddingstates.tree2einsum.py&grammar_ext.py: The core NLP-to-Quantum compiler. Translates Bobcat CCG derivation trees into parameterized unitary ansätze (e.g.,Sim14Ansatz,IQPAnsatz,BrickworkAnsatz).model.py: Contains theEinsumModelbuilt on PyTorch andcotengra. Compiles VQCs into optimizedeinsumcontraction paths for highly scalable, batched quantum simulation. Also contains the QInfoNCE loss variants.util.py: Implements fast, batched quantum gate operations (e.g.,BatchRz,BatchCRx) and the Fubini-Study distance metrics.visualisation.py: A comprehensive suite for generating PCA-projected 3D loss landscapes, t-SNE alignments, and fidelity density distributions usingplotlyandseaborn.example_training.ipynb: A complete end-to-end training and evaluation loop leveragingMLflowfor hyperparameter tracking.
We replace classical grammatical function applications with native quantum operations. Words are parameterized by ansätze acting on the
Figure 2: The structural compilation of "Alice likes Bob". Inductive grammatical biases are explicitly encoded into the circuit topology.
Standard quantum fidelity creates sharp loss landscapes prone to barren plateaus. To align text and image states contrastively, we employ a smooth Fubini-Study similarity metric:
Implemented in model.py as QInfoNCE_cos, this allows for highly stable gradient descent during multimodal alignment.
Figure 3: 3D projection of the parameter loss landscape during training, mapped via PCA.
QuLIP successfully bypasses the "bag-of-words" collapse seen in classical and unstructured quantum models.
| Model | Parameters | SVO-Swap | ARO Attribution | ARO Relation |
|---|---|---|---|---|
| QBoW | 100K | 50.00% | 50.00% | 50.00% |
| MicroCLIP | 100K | 68.42% | 50.85% | 51.05% |
| CLIP | 63M | 57.89% | 61.00% | 51.53% |
| QuLIP (CCG-VQC) | ~90K | 83.16% | 71.19% | 57.33% |
Ensure you have Python 3.12+ installed. The required dependencies include torch, lambeq, pennylane, cotengra, and mlflow.
# Clone the repository
git clone [https://github.com/YOUR_USERNAME/QuLIP.git](https://github.com/YOUR_USERNAME/QuLIP.git)
cd QuLIP
# Install dependencies (uv recommended for speed)
pip install uv
uv pip install lambeq pandas tqdm pennylane torch clip mlflow cotengra optuna
uv pip install git+[https://github.com/openai/CLIP.git](https://github.com/openai/CLIP.git)If you use this codebase or find our work on quantum compositional semantics helpful, please cite our paper:
@inproceedings{limbackstokin2026meaning,
title={Meaning Representations as Variational Quantum Circuits},
author={Limb\"{a}ck-Stokin, Tilen G. and Birdavade, Tanishka A. and Lo, Kin Ian and Sadrzadeh, Mehrnoosh},
booktitle={LREC},
year={2026},
organization={Quantum Learning Labs, University College London}
}
For questions or collaborations, please reach out:
- Tilen G. Limbäck-Stokin: tilen.limback-stokin.21@ucl.ac.uk
- Lab: Quantum Learning Labs, UCL
This project builds upon foundational work in Category Theory, Quantum Machine Learning, and Computational Linguistics. For the complete bibliography, see ref(1).bib.
- lambeq: An Efficient High-Level Python Library for Quantum NLP Kartsaklis et al. (2021). arXiv:2110.04236
- A CCG-Based Version of the DisCoCat Framework Yeung & Kartsaklis (2021). arXiv:2105.07720
- QNLP in Practice: Running Compositional Models of Meaning on a Quantum Computer Lorenz et al. (2023). arXiv:2102.12846
- Mathematical Foundations for a Compositional Distributional Model of Meaning Coecke, Sadrzadeh, & Clark (2010). arXiv:1003.4394
- Learning Transferable Visual Models From Natural Language Supervision (CLIP) Radford et al. (2021). arXiv:2103.00020
- When and why Vision-Language Models behave like Bags-of-Words, and what to do about it? Yuksekgonul et al. (2023). arXiv:2210.01936
- DisCoCLIP: A Distributional Compositional Tensor Network Encoder for Vision-Language Understanding Lo et al. (2025). arXiv:2509.21287
- Does CLIP Bind Concepts? Probing Compositionality in Large Image Models Lewis et al. (2024). arXiv:2212.10537
- Quantum Machine Learning Biamonte et al. (2017).. arXiv:1611.09347
- The Power of Quantum Neural Networks Abbas et al. (2021). arXiv:2011.00027
- Cost Function Dependent Barren Plateaus in Shallow Parametrized Quantum Circuits Cerezo et al. (2021). arXiv:2001.00550
- Expressibility and Entangling Capability of Parameterized Quantum Circuits Sim, Johnson, & Aspuru-Guzik (2019). arXiv:1905.10876