Snehal Raj

2026

1. arXiv

   “Train classical, deploy quantum” requires rethinking generalization

   Snehal Raj, Natansh Mathur, and Alejandro Perdomo-Ortiz

   arXiv preprint arXiv:2608.31117, 2026

   Abs Bib HTML Blog

   A leading strategy for quantum generative models trains the circuit on a classical computer with a moment-matching loss such as the MMD^2 and reserves the quantum device for sampling at deployment. We ask whether minimizing such a loss yields a model that generalizes. We prove that a moment loss of bounded capacity admits exact minimizers of exponentially small coverage, and we benchmark thirteen quantum and classical generative models by direct sampling on a cardinality-constrained dataset at up to 30 qubits and on genomic single-nucleotide variants. Models trained with a moment-matching loss generally generalize worse than likelihood-trained models, so a converged moment-matching loss is not a reliable measure of generalization.

   @article{raj2026tcdq,
     title = {``Train classical, deploy quantum'' requires rethinking generalization},
     author = {Raj, Snehal and Mathur, Natansh and Perdomo-Ortiz, Alejandro},
     journal = {arXiv preprint arXiv:2608.31117},
     year = {2026},
   }

2. arXiv

   Scalable Message-Passing Quantum Graph Neural Networks in the Weisfeiler–Leman Hierarchy

   Snehal Raj, Brian Coyle, Léo Monbroussou, and 3 more authors

   arXiv preprint arXiv:2606.26873, 2026

   Abs Bib HTML

   We show how a quantum graph neural network can be built to perform message passing, to be permutation equivariant, and to sit at a chosen level of the Weisfeiler–Leman hierarchy, the standard measure of how finely a model can tell graphs apart. Confining the dynamics to a particle-number subspace keeps the model trainable, and, as for classical graph networks, we pre-train on small graphs and deploy on larger ones with a readout whose cost stays low as the graph grows. We validate the framework in large-scale simulations of up to 56 qubits, on synthetic graphs that ordinary message passing cannot separate, on molecular property prediction, and on the travelling salesman problem.

   @article{raj2026qgnn,
     title = {Scalable Message-Passing Quantum Graph Neural Networks in the Weisfeiler--Leman Hierarchy},
     author = {Raj, Snehal and Coyle, Brian and Monbroussou, L{\'e}o and Ferreira-Martins, Andr{\'e} J. and Farias, Renato M. S. and Kashefi, Elham},
     journal = {arXiv preprint arXiv:2606.26873},
     year = {2026},
   }

3. arXiv

   Adaptive Directional Gradients for Parameterised Quantum Circuits

   Brian Coyle, Snehal Raj, Virag Umathe, and 2 more authors

   arXiv preprint arXiv:2606.09734, 2026

   Abs Bib HTML

   Training parameterised quantum circuits on hardware is bottlenecked by the measurement cost of gradient estimation, which under the parameter-shift rule scales linearly in the number of parameters. We propose a framework of forward gradient estimators that gives an unbiased gradient by averaging a tunable number of random directional derivatives, recovering SPSA, random coordinate descent, and the parameter-shift rule as limiting cases with no ancilla or controlled-gate overhead, and derive QUIVER, an adaptive measurement-frugal optimiser. Numerically, forward gradients train Hamming-weight-preserving orthogonal quantum neural networks with up to 60 qubits orders of magnitude more efficiently than the parameter-shift rule.

   @article{coyle2026adaptive,
     title = {Adaptive Directional Gradients for Parameterised Quantum Circuits},
     author = {Coyle, Brian and Raj, Snehal and Umathe, Virag and Cherrat, El Amine and Kashefi, Elham},
     journal = {arXiv preprint arXiv:2606.09734},
     year = {2026},
   }

2025

1. npj QI

   Training-Efficient Density Quantum Machine Learning

   Brian Coyle, Snehal Raj, Natansh Mathur, and 4 more authors

   npj Quantum Information, 2025

   Abs Bib HTML

   We introduce a novel approach to quantum machine learning that leverages density matrices for more efficient training. Our method demonstrates improved sample complexity and training efficiency compared to traditional quantum neural network approaches.

   @article{coyle2025training,
     title = {Training-Efficient Density Quantum Machine Learning},
     author = {Coyle, Brian and Raj, Snehal and Mathur, Natansh and Cherrat, El Amine and Jain, Nishant and Kazdaghli, Sofiene and Kerenidis, Iordanis},
     journal = {npj Quantum Information},
     volume = {11},
     number = {1},
     pages = {172},
     year = {2025},
     publisher = {Nature Publishing Group},
   }

2. arXiv

   QuIC: Quantum-Inspired Compound Adapters for Parameter Efficient Fine-Tuning

   Snehal Raj, and Brian Coyle

   arXiv preprint arXiv:2502.06916, 2025

   Abs Bib HTML

   We introduce QuIC, quantum-inspired adapter modules that operate with less than 0.02% memory footprint while maintaining pretrained representations through orthogonal weight parameters. Evaluated on LLaMA and vision transformers, our advanced configurations achieve over 40x compression compared to LoRA.

   @article{raj2025quic,
     title = {QuIC: Quantum-Inspired Compound Adapters for Parameter Efficient Fine-Tuning},
     author = {Raj, Snehal and Coyle, Brian},
     journal = {arXiv preprint arXiv:2502.06916},
     year = {2025},
   }

3. arXiv

   Bayesian Quantum Orthogonal Neural Networks for Anomaly Detection

   Natansh Mathur, Brian Coyle, Nishant Jain, and 4 more authors

   arXiv preprint arXiv:2504.18103, 2025

   Abs Bib HTML

   We propose a Bayesian approach to quantum orthogonal neural networks for anomaly detection tasks. Our method combines the expressiveness of quantum neural networks with Bayesian uncertainty quantification.

   @article{mathur2025bayesian,
     title = {Bayesian Quantum Orthogonal Neural Networks for Anomaly Detection},
     author = {Mathur, Natansh and Coyle, Brian and Jain, Nishant and Raj, Snehal and Tandon, Akshat and Krauser, Jasper Simon and Stoessel, Rainer},
     journal = {arXiv preprint arXiv:2504.18103},
     year = {2025},
   }

2023

1. Quantum

   Quantum Deep Hedging

   El Amine Cherrat, Snehal Raj, Iordanis Kerenidis, and 12 more authors

   Quantum, 2023

   Abs Bib HTML

   Quantum machine learning techniques are being explored for a variety of financial applications. In this work, we develop a quantum approach for the problem of deep hedging, a machine learning approach for managing the risk associated with derivatives trading.

   @article{cherrat2023quantum,
     title = {Quantum Deep Hedging},
     author = {Cherrat, El Amine and Raj, Snehal and Kerenidis, Iordanis and Shekhar, Abhishek and Wood, Ben and Dee, Jon and Chakrabarti, Shouvanik and Chen, Richard and Herman, Dylan and Hu, Shaohan and Minssen, Pierre and Shaydulin, Ruslan and Sun, Yue and Yalovetzky, Romina and Pistoia, Marco},
     journal = {Quantum},
     volume = {7},
     pages = {1191},
     year = {2023},
     doi = {10.22331/q-2023-11-29-1191},
   }

2021

1. EMNLP

   AUTOSUMM: Automatic Model Creation for Text Summarization

   Sharmila Reddy Nangi, Atharv Tyagi, Jay Mundra, and 4 more authors

   In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, 2021

   Abs Bib HTML

   We present AutoSumm, a system for automatic model creation for text summarization. Our approach enables the automated development of customized summarization models without extensive manual intervention.

   @inproceedings{nangi2021autosumm,
     title = {AUTOSUMM: Automatic Model Creation for Text Summarization},
     author = {Nangi, Sharmila Reddy and Tyagi, Atharv and Mundra, Jay and Mukherjee, Sagnik and Snehal, Raj and Chhaya, Niyati and Garimella, Aparna},
     booktitle = {Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing},
     pages = {9981--9991},
     year = {2021},
     organization = {Association for Computational Linguistics},
   }