Quantum Large Language Model Fine-Tuning

Paper Information

Title: Quantum Large Language Model Fine-Tuning

Authors: Sang Hyub Kim, Jonathan Mei, Claudio Girotto, Masako Yamada, Marin Roetteler

Published: eprint arXiv,(2025)

Reproduction Status: 🚧 In Progress

Reproducer: Cassandre Notton (cassandre.notton@quandela.com)

Abstract

This paper explores the use of a quantum parametrized circuit for LLM fine-tuning. The authors observe up to 3.14% improvements in accuracy over classical architectures of comparable model size.

Significance

This paper introduces a novel hybrid approach to LLM fine-tuning that addresses current limitations in the field. Given the critical role of fine-tuning in optimizing large language models for specific tasks, investigating how hybrid methodologies can enhance this process represents a significant research opportunity

Key Contributions Reproduced

We observe up to 3.14% improvements in accuracy over classical architectures

We present below the final results

Performance Comparison

Implementation Details

For a first analysis, we use a Generic Interferometer:

import merlin as ML # Package: merlinquantum, import: merlin
import numpy as np
import torch

device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')

builder = ML.CircuitBuilder(n_modes=modes)
builder.add_entangling_layer(trainable=True)
builder.add_angle_encoding(
    modes=list(range(X_train.shape[1])),
    scale=np.pi,
)
builder.add_entangling_layer(trainable=True)

# Create a simple quantum layer
q_layer = ML.QuantumLayer(
    input_size=X_train.shape[1],
    builder=builder,
    n_photons=modes // 2,
    measurement_strategy=ML.MeasurementStrategy.probs(),
)

Experimental Results

Model Performance Results
Method	Validation	Test
Classical Methods
LogisticRegression	0.8000	0.7669
SVC	0.8080	0.7846
MLP	0.8040	0.7701
SVC (paper)	0.8508	–
MLP (paper)	91.44	–
Quantum Methods
2 modes	0.5800	0.5691
4 modes	0.8120	0.7990
6 modes	0.8360	0.8280
8 modes	0.8480	0.8376
Single sQE 10Q (paper)	90.21	–
Multi sQE 14Q (paper)	92.7	–

We use smaller models as in the paper and are able to reach +4% in accuracy for the 8-mode interferometer. Next results will include a comparison of the number of parameters !

Interactive Exploration

Jupyter Notebook: ../notebooks/QLLM_fine_tuning

Citation

 @article{hyub2025quantum,
title={Quantum Large Language Model Fine-Tuning},
author={Hyub Kim, Sang and Mei, Jonathan and Girotto, Claudio and Yamada, Masako and Roetteler, Martin},
journal={arXiv e-prints},
pages={arXiv--2504},
year={2025}}

Future Work

Future Work includes:

thorough comparison of the performances with respect to the number of parameters;
analysis of the effect of the number of photons
experiments on SetFit using more than 2 classes for more complex classification

Note

End of document.