MerLin Reproduced Papers

About this repository

This repository contains implementations and resources for reproducing key quantum machine learning papers, with a focus on photonic and optical quantum computing.

It is part of the main MerLin project: https://github.com/merlinquantum/merlin and complements the online documentation available at:

https://merlinquantum.ai/research/reproduced_papers.html

Each paper reproduction is designed to be accessible, well-documented, and easy to extend. Contributions are welcome!

Papers reproduced:

Paper	Reproduction
Quantum Optical Reservoir Computing. Sources: munro_2024, rambach_2025, sakurai_simple_2025, lau_2025, sakurai_2025	- Scalability: increasing number of modes leads to better performance for the reservoir on MNIST. - In the original work and in our reproduction, we see a quantum boost with modest resources. - Our reproduction is QPU compliant.
Computational Advantage in Hybrid Quantum NeuralNetworks: Myth or Reality?. Source: kashif_2024	The original paper and our reproduction display that an HQNN model requires less parameters than a classical NN to achieve at least 90% accuracy on the noisy spiral dataset used that has a variable number of features (between 5 and 60).
Quantum Self-Supervised Learning. Source: jaderberg_2021	The MerLin model is better and faster than qiskit. On the first five classes of CIFAR-10 with two epochs and only eight modes for the QSSL: - $\times0.97$ speedup versus a fully classical model compared to qiskit's $\times0.08$ speedup. - Accuracy of 49.2% compared to 48.4% for qiskit and 48.1% for a classical SSL.
Large-Language Model Fine-Tuning. Source: kim_2025	The original paper states that the quantum enhanced model improves the accuracy to up to 3.14% compared to classical models with comparable number of parameters on a sentiment classification task on text data. On our end, all the best performing models (whether quantum or classical) reach around 89% accuracy without a clear segmentation.
Quantum Long Short-Term Memory. Source: chen_2020	Our MerLin-based photonic QLSTM yields similar results to the original gate-based QLSTM on function fitting tasks. However, the weaknesses of the classical LSTM reported in the paper were not fully present in our reproduction.
Fock State-enhanced expressivity of Quantum Machine Learning Models. Source: gan_2022	As explained and displayed in the original paper, our experiments also showcase that an increase in the number of photons used (when using the data encoding scheme that is proposed) is intrinsically linked to an increase in Variational Quantum Circuit (VQC) expressivity .
Photonic Quantum Convolutional Neural Networks with Adaptive State Injection. Source: monbroussou_2025	Our reproduction improved some of the reported accuracies for binary image classification using the proposed model by optimizing hyperparameters: - Test accuracy on Custom BAS went from 92.7 ± 2.1 % to 98.2 ± 2.2 %. - Test accuracy on MNIST (0 vs 1) went from 93.1 ± 3.6 % to 98.8 ± 1.0 %.
Quantum Convolutional Neural Networks. Source hur_2022	The source paper reports that under similar training budgets, the QCNN outperforms the CNN. The results were reproduced on a photonic architecture, however in such architecture, we conclude that classical models remain more parameter-efficient, but they cap out in absolute accuracy.
Quantum Relational Knowledge Distillation. Source: liu_2025	In both the reference paper and in our reproduction, we see that the improvement of the student model due to the distillation is superior in the quantum relational knowledge distillation scheme compared to in its classical counterpart.
Quantum Recurrent Neural Networks for Sequential Learning. Source: li_2023	(To complete when reproduction is done).
Distributed Quantum Neural Networks on Distributed Photonic Quantum Computing. Source: chen_2025	The paper and our reproduction reach the conclusion that fewer quantum parameters need to be trained to obtain all the classical parameters. We also reach these following results: - Accuracy better or worse with an approximate 2% error for bond dimensions in the training accuracy. - Accuracy better or worse with an approximate 4% error for bond dimensions in the testing accuracy with 4 times less epochs. - Speedup in the training using ADAM optimizer instead of COBYLA. Nonetheless, we attain different results for the ablation study.
Data Reuploading mauser_2025	Our results confirm that the fully quantum data reuploading model is well performing and resource-efficient in the context of binary classification on the four datasets used. We also obtain that the model's expressivity scales with its number of reuplading layers.
Nearest Centroids. Source: johri_2020	Reproducing this photon native algorithm led to accuracies that match the ones obtained classically on the three datasets of interest, as reported in the source paper.
Photonic QGAN. Source sedrakyan_2024	(To complete when reproduction is done).
Quantum Enhanced Kernels. Source: yin_2025	(To complete when reproduction is done).
Quantum Transfer Learning. Source: mari_2020	Three transfer learning frameworks are reproduced: classical to classical, classical to quantum and quantum to classical. We obtain about the same accuracy with the MerLin implementation and the simulated runs of the paper (in a classical to quantum learning experiment). Photon count seems non-influential in this specific setting.
Adversarial Learning. Source: lu_2020	We observed on a photonic model, just like the authors of the paper on a gate-based model, that quantum classifiers are vulnerable to direct and transferred adversarial attacks but adversarial training is also effective against specific attack types. MNIST classification (1 vs 9): - Clean accuracy = 98% - Adversarial accuracy (BIM, $\epsilon=0.1$) = 15% - Adversarial accuracy post-adversarial training (BIM) = 95%
Photonic Quantum Memristor. Source: selimovic_2025	We developed a photonic quantum reservoir computing architecture with the addition of a quantum memristor which acts as a feedback loop (memory). We noticed that in both cases the use of the memristor enhances the non-linear capabilities of the model and this leads to improved performance compared to the non-memristor case and some classical benchmarks. The results are similar to those presented in the paper and the error is 5 times smaller on the learning task of the NARMA dataset.

Running existing reproductions

• Browse the up-to-date catalogue at https://merlinquantum.ai/reproduced_papers/index.html to pick the paper you want to execute. Every paper now lives under papers/<NAME>/; the <NAME> you pass to the CLI is just that folder name (e.g., QLSTM, QORC, reproduction_template).

You can also list available reproductions with python implementation.py --list-papers.

• cd into papers/<NAME> and install its dependencies: pip install -r requirements.txt (each reproduction keeps its own list).

• Launch training/eval runs through the shared CLI from the repo root (the runner will cd into the project automatically):

   python implementation.py --paper <NAME> --config configs/<config>.json

• If you prefer running from inside papers/<NAME>, reference the repo-level runner: python ../../implementation.py --config configs/<config>.json (no --paper flag needed when executed from within the project).

All logs, checkpoints, and figures land in papers/<NAME>/outdir/run_YYYYMMDD-HHMMSS/ unless the configs specify a different base path.

Need a quick tour of a project’s knobs? Run python implementation.py --paper <NAME> --help to print the runtime-generated CLI for that reproduction (dataset switches, figure toggles, etc.) before launching a full experiment.

Data location

• Default data root is data/ at the repo root; each paper writes under data/<NAME>/ to avoid per-venv caches.
• Override with DATA_DIR=/abs/path or python implementation.py --data-root /abs/path ... (applies to the current run and is exported to downstream loaders).

Shared data helpers:

• Common dataset-generation code lives under papers/shared/<paper>/ when multiple reproductions reuse the same logic. Each paper exposes a thin lib/data.py (or equivalent) that simply imports from the shared module.
• If you add new shared data utilities, place them in papers/shared/<paper>/ and have paper-local lib/ importers forward to them so tests and runners stay stable.

Universal CLI flags provided by the shared runner:

• --seed INT Reproducibility seed propagated to Python/NumPy/PyTorch backends.
• --dtype STR Force a global tensor dtype before model-specific overrides.
• --device STR Torch device string (cpu, cuda:0, mps, ...).
• --log-level LEVEL Runtime logging verbosity (INFO by default). Project-specific cli.json files only declare the extra paper knobs; the runner injects the global options automatically.

Smoke-test all papers quickly

• From repo root, run the portable smoke harness scripts/smoke_test_all_papers.sh to install per-paper venvs under .smoke_envs/, execute each paper’s default config, and run pytest, logging to .smoke_logs/<paper>.log.
• Pass an optional substring to target specific papers (faster dev loop): scripts/smoke_test_all_papers.sh QRKD only runs papers whose names contain QRKD.
• Timeout markers appear in logs when a run or test exceeds the limit; rerun after adjusting configs or deps as needed.

Precision control (dtype)

• Every reproduction accepts an optional top-level "dtype" entry in its configs, just like "seed". When present, the shared runner casts input tensors and initializes models in that dtype.
• Individual models can still override via model.dtype; if omitted, each reproduction picks a sensible default (e.g., float64 for photonic MerLin layers).
• Use this to downgrade to float32 for speed, experiment with bfloat16, or enforce float64 reproducibility across classical/quantum variants.

How to contribute a reproduced paper

We encourage contributions of new quantum ML paper reproductions. Please follow the guidelines in the how_to_contribute file