arxiv: 2602.11092 · v2 · submitted 2026-02-11 · 💻 cs.LG · cs.PL· quant-ph

Recognition: 2 theorem links

· Lean Theorem

MerLin: A Discovery Engine for Photonic and Hybrid Quantum Machine Learning

Cassandre Notton , Benjamin Stott , Philippe Schoeb , Anthony Walsh , Gr\'egoire Leboucher , Vincent Espitalier , Vassilis Apostolou , Louis-F\'elix Vigneux

show 2 more authors

Alexia Salavrakos Jean Senellart

Authors on Pith no claims yet

Pith reviewed 2026-05-16 02:23 UTC · model grok-4.3

classification 💻 cs.LG cs.PLquant-ph

keywords photonic quantum machine learninghybrid QMLlinear optical circuitsdifferentiable simulationreproducibilitybenchmarkingPyTorch integrationquantum layers

0 comments

The pith

MerLin adds optimized simulation of linear optical circuits to PyTorch so photonic quantum layers can be trained end-to-end with standard machine learning tools.

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

The paper introduces MerLin as an open-source framework that embeds fast simulation of linear optical circuits directly into PyTorch and scikit-learn workflows. This setup supports end-to-end differentiable training of quantum layers in photonic and hybrid quantum machine learning models. The authors demonstrate the approach by reproducing results from eighteen state-of-the-art papers across kernel methods, reservoir computing, convolutional and recurrent networks, generative models, and modern training methods. A sympathetic reader would value the result because it turns scattered algorithmic proposals into a shared, extensible platform that works with familiar tooling and can test against real hardware constraints.

Core claim

MerLin is an open-source framework that integrates optimized strong simulation of linear optical circuits into PyTorch and scikit-learn, enabling end-to-end differentiable training of quantum layers. The authors use it to reproduce eighteen state-of-the-art photonic and hybrid QML works spanning kernel methods, reservoir computing, convolutional and recurrent architectures, generative models, and modern training paradigms. These reproductions are released as reusable modular experiments to establish a shared experimental baseline.

What carries the argument

Optimized strong simulation of linear optical circuits integrated into PyTorch and scikit-learn for end-to-end differentiable training of quantum layers.

If this is right

Researchers can apply existing ablation studies, cross-modality comparisons, and hybrid classical-quantum workflows using standard machine learning libraries.
The framework supports direct tests on available quantum hardware while allowing exploration beyond current hardware limits.
The released modular experiments create a consistent baseline for systematic benchmarking across photonic QML approaches.
MerLin functions as a co-design tool that links algorithm development with hardware constraints.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Standardizing photonic QML experiments in this way could make it simpler to run controlled comparisons that reveal where quantum elements improve performance on specific tasks.
Community extensions of the modular experiments might quickly test new datasets or training methods without rebuilding the simulation layer from scratch.
The same framework could later incorporate other quantum modalities once their simulators become available, enabling broader hybrid studies.

Load-bearing premise

The reproduced experiments match the original papers and the simulation captures the physical behavior that actually determines machine learning performance on real hardware.

What would settle it

Running one of the reproduced experiments on actual photonic hardware and obtaining performance metrics that differ substantially from the simulated results reported in MerLin.

Figures

Figures reproduced from arXiv: 2602.11092 by Alexia Salavrakos, Anthony Walsh, Benjamin Stott, Cassandre Notton, Gr\'egoire Leboucher, Jean Senellart, Louis-F\'elix Vigneux, Philippe Schoeb, Vassilis Apostolou, Vincent Espitalier.

**Figure 1.** Figure 1: MerLin architecture for photonic QML. (A) Classical data encoding and photonic circuit configuration define the quantum model. (B) MerLin integrates PyTorch-based optimization with photonic-native execution through a logical-tophotonic bridge, differentiable quantum layers, and hardware-oriented simulation, with optional inference on Quandela photonic QPUs. (C) Measurement strategies and detectors behavio… view at source ↗

**Figure 2.** Figure 2: Overview of a photonic VQC implemented with [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

read the original abstract

Identifying where quantum models may offer practical benefits in near term quantum machine learning (QML) requires moving beyond isolated algorithmic proposals toward systematic and empirical exploration across models, datasets, and hardware constraints. We introduce MerLin, an open-source framework designed as a discovery engine for photonic and hybrid quantum machine learning. MerLin integrates optimized strong simulation of linear optical circuits into standard PyTorch and scikit learn workflows, enabling end-to-end differentiable training of quantum layers. MerLin is designed around systematic benchmarking and reproducibility. As an initial contribution, we reproduce eighteen state-of-the-art photonic and hybrid QML works spanning kernel methods, reservoir computing, convolutional and recurrent architectures, generative models, and modern training paradigms. These reproductions are released as reusable, modular experiments that can be directly extended and adapted, establishing a shared experimental baseline consistent with empirical benchmarking methodologies widely adopted in modern artificial intelligence. By embedding photonic quantum models within established machine learning ecosystems, MerLin allows practitioners to leverage existing tooling for ablation studies, cross-modality comparisons, and hybrid classical-quantum workflows. The framework already implements hardware-aware features, allowing tests on available quantum hardware while enabling exploration beyond its current capabilities, positioning MerLin as a forward-looking co-design tool linking algorithms, benchmarks, and hardware.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

MerLin is a practical open-source framework that wires photonic circuit simulation into PyTorch and scikit-learn for differentiable training and ships 18 modular reproductions as shared baselines.

read the letter

MerLin wires optimized linear optical circuit simulation into standard PyTorch and scikit-learn pipelines so quantum layers can be trained end-to-end with backprop. The main deliverable is the released set of 18 modular reproductions covering kernels, reservoirs, convolutional and recurrent models, and generative approaches. That gives the field a concrete starting point for comparisons instead of scattered one-off implementations. The hardware-aware hooks are a sensible addition for co-design work. The paper does the engineering cleanly and avoids overclaiming new algorithms. The soft spots are narrow. The whole argument rests on the reproductions being faithful to the originals and the simulator capturing the physics that actually matters for learning performance. A referee would want explicit side-by-side metrics or error tables showing how close each reproduction lands. If those checks hold, the contribution is solid. This is for people already working in near-term photonic or hybrid QML who need reusable code rather than a new theoretical result. A reader who wants to run ablations or extend existing models will get immediate value from the modular experiments. It deserves a serious referee because the reproducibility infrastructure is the point and the implementation details need checking. I would send it to review.

Referee Report

1 major / 2 minor

Summary. The manuscript introduces MerLin, an open-source framework integrating optimized strong simulation of linear optical circuits into PyTorch and scikit-learn for end-to-end differentiable training of quantum layers in photonic and hybrid QML. It reproduces eighteen state-of-the-art works across kernel methods, reservoir computing, convolutional/recurrent architectures, generative models, and training paradigms, releasing them as modular, reusable experiments to establish shared baselines.

Significance. If the simulations accurately model relevant photonic behaviors under hardware constraints and the reproductions faithfully match the originals, MerLin would provide a valuable discovery engine and benchmarking platform. It embeds quantum models in standard ML workflows, supports ablation studies and hybrid co-design, and aligns with empirical reproducibility standards in AI, potentially accelerating systematic exploration in near-term photonic QML.

major comments (1)

The central claim of establishing consistent baselines via 18 reproductions is load-bearing, yet the manuscript provides insufficient quantitative verification details (e.g., direct metric comparisons or deviation analysis between reproduced and original results) to confirm fidelity, particularly given the simulation approximations inherent to strong linear optical circuit modeling.

minor comments (2)

Clarify in the implementation section how the strong simulation optimizations scale with circuit size and photon number, including any trade-offs for differentiability in PyTorch workflows.
The abstract and introduction would benefit from explicit mention of the specific hardware constraints (e.g., loss, noise models) already implemented, to better frame the forward-looking co-design claims.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the positive assessment and the recommendation for minor revision. We address the single major comment below and will strengthen the manuscript accordingly.

read point-by-point responses

Referee: The central claim of establishing consistent baselines via 18 reproductions is load-bearing, yet the manuscript provides insufficient quantitative verification details (e.g., direct metric comparisons or deviation analysis between reproduced and original results) to confirm fidelity, particularly given the simulation approximations inherent to strong linear optical circuit modeling.

Authors: We agree that explicit quantitative verification strengthens the reproducibility claim. While the manuscript and accompanying repository release modular reproduction scripts that allow exact replication of each experiment, we acknowledge that the main text would benefit from direct side-by-side metric comparisons and deviation statistics. In the revised version we will add a dedicated appendix (or expanded results section) containing tables that report original versus reproduced performance metrics for all 18 works, including absolute and relative deviations. These tables will also note any sources of discrepancy arising from the strong-simulation model (e.g., floating-point precision or hardware-aware approximations), thereby confirming fidelity under the modeling assumptions used. revision: yes

Circularity Check

0 steps flagged

No significant circularity in framework and reproduction claims

full rationale

The manuscript presents MerLin as an open-source simulation and benchmarking framework that integrates linear optical circuit simulation into PyTorch/scikit-learn workflows and releases modular reproductions of 18 prior photonic/hybrid QML experiments. No derivation chain, fitted parameter, or uniqueness theorem is claimed; the central contributions are software integration and empirical baselines. These do not reduce to self-definition, self-citation load-bearing, or renaming of fitted results by construction. The paper is self-contained against external benchmarks and code release.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

MerLin is a software framework contribution. It relies on standard assumptions from quantum optics for linear optical circuit simulation and on the correctness of PyTorch autograd for custom layers. No new physical entities or fitted parameters are introduced in the abstract.

axioms (1)

domain assumption Strong simulation of linear optical circuits can be made end-to-end differentiable and integrated into standard ML training loops.
Invoked in the description of enabling differentiable training of quantum layers.

pith-pipeline@v0.9.0 · 5570 in / 1303 out tokens · 73944 ms · 2026-05-16T02:23:48.404090+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

MerLin integrates optimized strong simulation of linear optical circuits into standard PyTorch and scikit-learn workflows, enabling end-to-end differentiable training of quantum layers... We reproduce eighteen state-of-the-art photonic and hybrid QML works
IndisputableMonolith/Foundation/RealityFromDistinction.lean reality_from_one_distinction unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

The state at the output of the interferometer is then measured by photon detectors... probability of measuring each of them is determined by the matrix permanent of submatrices of U

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Forward citations

Cited by 1 Pith paper

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

Pre-Asymptotic Trainability in Photonic Variational Circuits under Postselection
quant-ph 2026-05 unverdicted novelty 6.0

Photonic variational circuits exhibit polynomial gradient decay in some postselection regimes and exponential decay in others, governed by postselection geometry rather than Hilbert-space dimension.

Reference graph

Works this paper leans on

48 extracted references · 48 canonical work pages · cited by 1 Pith paper · 3 internal anchors

[1]

DeepQuantum: A PyTorch-based Software Platform for Quantum Machine Learning and Photonic Quantum Computing

J.-J. Heet al., “DeepQuantum: A PyTorch-based Software Platform for Quantum Machine Learning and Photonic Quantum Computing,” Dec. 2025, arXiv:2512.18995

work page internal anchor Pith review Pith/arXiv arXiv 2025
[2]

Perceval: A Software Platform for Discrete Variable Photonic Quantum Computing,

N. Heurtelet al., “Perceval: A Software Platform for Discrete Variable Photonic Quantum Computing,”Quantum, vol. 7, p. 931, Feb. 2023

work page 2023
[3]

A versatile single-photon-based quantum computing platform,

N. Maringet al., “A versatile single-photon-based quantum computing platform,”Nature Photonics, vol. 18, no. 6, pp. 603–609, Jun. 2024

work page 2024
[4]

Quantum energetic advantage before computational advantage in boson sampling,

A. Soretet al., “Quantum energetic advantage before computational advantage in boson sampling,”arXiv preprint arXiv:2601.08068, 2026

work page arXiv 2026
[5]

Quantum computing with qiskit,

A. Javadi-Abhariet al., “Quantum computing with qiskit,” 2024

work page 2024
[6]

Cirq Developers, “Cirq,” https://github.com/quantumlib/Cirq, 2021

work page 2021
[7]

Pulser: An open-source package for the design of pulse sequences in programmable neutral-atom arrays,

H. Silv ´erioet al., “Pulser: An open-source package for the design of pulse sequences in programmable neutral-atom arrays,”Quantum, vol. 6, p. 629, 2022

work page 2022
[8]

Strawberry Fields: A Software Platform for Photonic Quantum Computing

N. Killoranet al., “Strawberry Fields: A Software Platform for Pho- tonic Quantum Computing,”Quantum, vol. 3, p. 129, Mar. 2019, arXiv:1804.03159

work page internal anchor Pith review Pith/arXiv arXiv 2019
[9]

Piquasso: A Photonic Quantum Computer Simulation Software Platform,

Z. Kolarovszkiet al., “Piquasso: A Photonic Quantum Computer Simulation Software Platform,”Quantum, vol. 9, p. 1708, Apr. 2025, arXiv:2403.04006

work page arXiv 2025
[10]

PennyLane: Automatic differentiation of hybrid quantum-classical computations

V . Bergholmet al., “Pennylane: Automatic differentiation of hybrid quantum-classical computations,” 2018, arXiv:1811.04968

work page internal anchor Pith review Pith/arXiv arXiv 2018
[11]

QuantumNAS: Noise-Adaptive Search for Robust Quantum Circuits,

H. Wanget al., “QuantumNAS: Noise-Adaptive Search for Robust Quantum Circuits,” in2022 IEEE International Symposium on High- Performance Computer Architecture (HPCA), Apr. 2022, pp. 692–708, arXiv:2107.10845

work page arXiv 2022
[12]

Qiskit-torch-module: Fast prototyping of quantum neural networks,

N. Meyeret al., “Qiskit-torch-module: Fast prototyping of quantum neural networks,” in2024 IEEE International Conference on Quantum Computing and Engineering (QCE), vol. 1. IEEE, 2024, pp. 817–823

work page 2024
[13]

Establishing baselines for photonic quantum machine learning: Insights from an open, collaborative initiative,

C. Nottonet al., “Establishing baselines for photonic quantum machine learning: Insights from an open, collaborative initiative,” 2025

work page 2025
[14]

Fock state-enhanced expressivity of quantum machine learning models,

B. Y . Ganet al., “Fock state-enhanced expressivity of quantum machine learning models,”EPJ Quantum Technology, vol. 9, no. 1, p. 16, 2022

work page 2022
[15]

Experimental quantum-enhanced kernel-based machine learning on a photonic processor,

Z. Yinet al., “Experimental quantum-enhanced kernel-based machine learning on a photonic processor,”Nature Photonics, vol. 19, no. 9, pp. 1020–1027, 2025

work page 2025
[16]

Discopy: Monoidal categories in python,

G. De Feliceet al., “Discopy: Monoidal categories in python,”arXiv preprint arXiv:2005.02975, 2020

work page arXiv 2005
[17]

Quantum optical reservoir computing powered by boson sampling,

A. Sakuraiet al., “Quantum optical reservoir computing powered by boson sampling,”Optica Quantum, vol. 3, no. 3, pp. 238–245, Jun. 2025

work page 2025
[18]

Photonic Quantum-Accelerated Machine Learning,

M. Rambachet al., “Photonic Quantum-Accelerated Machine Learning,” Dec. 2025, arXiv:2512.08318

work page arXiv 2025
[19]

Photonic quantum convolutional neural net- works with adaptive state injection,

L. Monbroussouet al., “Photonic quantum convolutional neural net- works with adaptive state injection,” 2025

work page 2025
[20]

Qoptcraft: A python package for the design and study of linear optical quantum systems,

D. G. Aguadoet al., “Qoptcraft: A python package for the design and study of linear optical quantum systems,”Computer Physics Communi- cations, vol. 282, p. 108511, 2023

work page 2023
[21]

Better than classical? the subtle art of benchmarking quantum machine learning models,

J. Bowleset al., “Better than classical? the subtle art of benchmarking quantum machine learning models,”arXiv preprint arXiv:2403.07059, 2024

work page arXiv 2024
[22]

Solving graph problems with single-photons and linear optics,

R. Mezheret al., “Solving graph problems with single-photons and linear optics,” Aug. 2023, arXiv:2301.09594 [quant-ph]

work page arXiv 2023
[23]

Strong simulation of linear optical processes,

N. Heurtelet al., “Strong simulation of linear optical processes,” Computer Physics Communications, vol. 291, p. 108848, Oct. 2023

work page 2023
[24]

Pytorch: An imperative style, high-performance deep learning library,

A. Paszkeet al., “Pytorch: An imperative style, high-performance deep learning library,” 2019

work page 2019
[25]

Quantum circuit compression using qubit logic on qudits,

L. Lysaghtet al., “Quantum circuit compression using qubit logic on qudits,” 2024

work page 2024
[26]

The effect of data encoding on the expressive power of variational quantum machine learning models,

M. Schuldet al., “The effect of data encoding on the expressive power of variational quantum machine learning models,”Physical Review A, vol. 103, no. 3, p. 032430, Mar. 2021, arXiv:2008.08605

work page arXiv 2021
[27]

Experimental neuromorphic computing based on quantum memristor,

M. Selimovi ´cet al., “Experimental neuromorphic computing based on quantum memristor,” 2025

work page 2025
[28]

Nearest centroid classification on a trapped ion quantum computer,

S. Johriet al., “Nearest centroid classification on a trapped ion quantum computer,” p. 122, 2021

work page 2021
[29]

Experimental data re-uploading with provable enhanced learning capabilities,

M. F. Mauseret al., “Experimental data re-uploading with provable enhanced learning capabilities,” 2025

work page 2025
[30]

Quantum Long Short-Term Memory,

S. Y .-C. Chenet al., “Quantum Long Short-Term Memory,” Sep. 2020, arXiv:2009.01783

work page arXiv 2020
[31]

Quantum Recurrent Neural Networks for Sequential Learning,

Y . Liet al., “Quantum Recurrent Neural Networks for Sequential Learning,” Feb. 2023, arXiv:2302.03244

work page arXiv 2023
[32]

Quantum convolutional neural network for classical data classification,

T. Huret al., “Quantum convolutional neural network for classical data classification,”Quantum Machine Intelligence, vol. 4, no. 1, p. 3, Jun. 2022, arXiv:2108.00661

work page arXiv 2022
[33]

Quantum Relational Knowledge Distillation,

C.-Y . Liuet al., “Quantum Relational Knowledge Distillation,” Aug. 2025, arXiv:2508.13054

work page arXiv 2025
[34]

Distributed quantum neural networks on distributed photonic quantum computing,

K.-C. Chenet al., “Distributed quantum neural networks on distributed photonic quantum computing,” IEEE, pp. 1477–1488, 2025

work page 2025
[35]

Quantum Self-Supervised Learning,

B. Jaderberget al., “Quantum Self-Supervised Learning,” 2021

work page 2021
[36]

Transfer learning in hybrid classical-quantum neural networks,

A. Mariet al., “Transfer learning in hybrid classical-quantum neural networks,”Quantum, vol. 4, p. 340, 2020

work page 2020
[37]

Photonic quantum generative adversarial networks for classical data,

T. Sedrakyanet al., “Photonic quantum generative adversarial networks for classical data,”Optica Quantum, vol. 2, no. 6, pp. 458–467, 2024

work page 2024
[38]

Computational Advantage in Hybrid Quantum Neural Networks: Myth or Reality?

M. Kashifet al., “Computational Advantage in Hybrid Quantum Neural Networks: Myth or Reality?” Dec. 2024, arXiv:2412.04991

work page arXiv 2024
[39]

Quantum Large Language Model Fine-Tuning,

S. H. Kimet al., “Quantum Large Language Model Fine-Tuning,” Apr. 2025, arXiv:2504.08732

work page arXiv 2025
[40]

Quantum adversarial machine learning,

S. Luet al., “Quantum adversarial machine learning,”Physical Review Research, vol. 2, no. 3, p. 033212, 2020

work page 2020
[41]

Quantum vs. classical: A comprehensive benchmark study for predicting time series with variational quantum machine learning,

T. Fellneret al., “Quantum vs. classical: A comprehensive benchmark study for predicting time series with variational quantum machine learning,”Machine Learning: Science and Technology, vol. 7, no. 1, p. 010501, 2026

work page 2026
[42]

Deep residual learning for image recognition,

K. Heet al., “Deep residual learning for image recognition,” pp. 770– 778, 2016

work page 2016
[43]

Subspace preserving quantum convolutional neural network architectures,

L. Monbroussouet al., “Subspace preserving quantum convolutional neural network architectures,”Quantum Science and Technology, vol. 10, no. 2, p. 025050, Mar. 2025

work page 2025
[44]

The “echo state

H. Jaeger, “The “echo state” approach to analysing and training recurrent neural networks,”GMD-Report 148, German National Research Institute for Computer Science, 01 2001

work page 2001
[45]

Real-time computing without stable states: A new framework for neural computation based on perturbations,

W. Maasset al., “Real-time computing without stable states: A new framework for neural computation based on perturbations,”Neural Computation, vol. 14, pp. 2531–2560, 11 2002

work page 2002
[46]

Dinov2: Learning robust visual features without supervision,

M. Oquabet al., “Dinov2: Learning robust visual features without supervision,” 2024

work page 2024
[47]

Quantum generative adversarial networks,

P.-L. Dallaire-Demers and N. Killoran, “Quantum generative adversarial networks,”Physical Review A, vol. 98, no. 1, jul 2018

work page 2018
[48]

An overview of the simultaneous perturbation method for efficient optimization,

J. C. Spall, “An overview of the simultaneous perturbation method for efficient optimization,”Johns Hopkins apl technical digest, vol. 19, no. 4, pp. 482–492, 1998

work page 1998