Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Quantum-Enhanced LLM Efficient Fine Tuning

Xiaofei Kong, Lei Li, Zhaoyun Chen, Cheng Xue, Xiaofan Xu, Huanyu Liu, Yuchun Wu, Yuan Fang, Han Fang, Kejiang Chen, Yang Yang, Menghan Dou, Guoping Guo

TL;DR

The paper tackles the expressivity bottleneck of classical PEFT methods like LoRA for large language models by proposing Quantum Tensor Hybrid Adaptation (QTHA), which fuses Matrix Product Operator tensor networks with a quantum neural network. The MPO reduces parameters while the QNN injects high-dimensional nonlinear features, combined through a linear fusion to form the updated representations. Empirical results show QTHA achieves up to 76% fewer trainable parameters and up to 17% improvements in training loss and test metrics across multiple datasets, with successful validation on quantum hardware backends such as Origin Wukong. This work demonstrates a practical pathway for quantum-enhanced fine-tuning of billion-parameter LLMs and lays groundwork for future quantum-assisted AI systems.

Abstract

Low-Rank Adaptation (LoRA) enables efficient fine-tuning of pre-trained language models through low-rank matrix approximation, achieving effectiveness in many scenarios. However, its representation capacity is constrained in complex tasks or high-rank dependency settings, potentially limiting model adaptability. To overcome the expressive bottleneck in classical low-rank approximation for fine-tuning large language models (LLMs), we propose Quantum Tensor Hybrid Adaptation (QTHA), a parameter-efficient fine-tuning method that integrates a quantum neural network (QNN) with a tensor network. QTHA explores quantum tensor hybrid fine-tuning within low-rank spaces by decomposing pre-trained weights into quantum neural network and tensor network representations, leveraging quantum state superposition to overcome classical rank limitations. Experiments demonstrate that QTHA achieves performance comparable to or surpassing LoRA in parameter-efficient fine-tuning. Compared to LoRA, QTHA reduces trainable parameters by 76% while reducing training loss by up to 17% and improving test set performance by up to 17% within the same training steps. This research not only enables lightweight adaptation of quantum resources to the billion-parameter models but also validates the feasibility of quantum hardware optimization driven by LLM tasks. It establishes the first engineering-ready foundation for future quantum-enhanced Artificial General Intelligence (AGI) systems.

Quantum-Enhanced LLM Efficient Fine Tuning

TL;DR

The paper tackles the expressivity bottleneck of classical PEFT methods like LoRA for large language models by proposing Quantum Tensor Hybrid Adaptation (QTHA), which fuses Matrix Product Operator tensor networks with a quantum neural network. The MPO reduces parameters while the QNN injects high-dimensional nonlinear features, combined through a linear fusion to form the updated representations. Empirical results show QTHA achieves up to 76% fewer trainable parameters and up to 17% improvements in training loss and test metrics across multiple datasets, with successful validation on quantum hardware backends such as Origin Wukong. This work demonstrates a practical pathway for quantum-enhanced fine-tuning of billion-parameter LLMs and lays groundwork for future quantum-assisted AI systems.

Abstract

Low-Rank Adaptation (LoRA) enables efficient fine-tuning of pre-trained language models through low-rank matrix approximation, achieving effectiveness in many scenarios. However, its representation capacity is constrained in complex tasks or high-rank dependency settings, potentially limiting model adaptability. To overcome the expressive bottleneck in classical low-rank approximation for fine-tuning large language models (LLMs), we propose Quantum Tensor Hybrid Adaptation (QTHA), a parameter-efficient fine-tuning method that integrates a quantum neural network (QNN) with a tensor network. QTHA explores quantum tensor hybrid fine-tuning within low-rank spaces by decomposing pre-trained weights into quantum neural network and tensor network representations, leveraging quantum state superposition to overcome classical rank limitations. Experiments demonstrate that QTHA achieves performance comparable to or surpassing LoRA in parameter-efficient fine-tuning. Compared to LoRA, QTHA reduces trainable parameters by 76% while reducing training loss by up to 17% and improving test set performance by up to 17% within the same training steps. This research not only enables lightweight adaptation of quantum resources to the billion-parameter models but also validates the feasibility of quantum hardware optimization driven by LLM tasks. It establishes the first engineering-ready foundation for future quantum-enhanced Artificial General Intelligence (AGI) systems.

Paper Structure

This paper contains 10 sections, 13 equations, 2 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Methods
  3. Quantum Tensor Network Based on MPO
  4. Quantum Tensor Hybrid Adaptation
  5. Experiment
  6. Datasets
  7. Evaluation Metrics
  8. Results
  9. Implementation on quantum hardware
  10. Conclusion

Figures (2)

  • Figure 1: Schematic diagram of the Quantum Tensor Hybrid Adapation (QTHA). QTHA dynamically adjusts feature weights through parameter tuning and outputs a combination of features from MPO and QNN.
  • Figure 2: Training loss comparison of QTHA, LoRA and QuanTA across varying sample sizes (N) and datasets. Subplots (a–b) illustrate results on the CPsyCounD dataset, while subplots (c–d) and (e–f) show corresponding analyses on the R1-Distill-SFT and CH-R1-Math dataset. QTHA consistently achieves lower training losses compared to baseline methods, demonstrating enhanced convergence efficiency under diverse experimental conditions.