LLM-QFL: Distilling Large Language Model for Quantum Federated Learning

TL;DR

The paper tackles inefficiencies in quantum federated learning by introducing LLM-QFL, a framework that distills and federated-fines-tunes large language models within QFL. It leverages federated distillation, where locally fine-tuned LLMs guide quantum model optimization and act as reinforcement agents to regulate optimizer steps, client selection, and early termination, backed by KL-based distillation and a global aggregation rule. The authors prove a convergence rate of $\mathcal{O}(\frac{1}{T})$ under standard assumptions and demonstrate practical gains in communication efficiency and training speed, aided by parameter-efficient tuning via LoRA/QLoRA and adaptive optimization. Empirical validation spans genomic (DemoHumanOrWorm) and language (TweetEval) tasks on IBM QPU and simulators, showing substantial idle computation reductions (~30%) and improved convergence, while highlighting real-hardware noise and queueing effects. The work offers a principled, scalable route to combine LLMs with quantum learning, with potential impact on privacy-preserving, efficient distributed quantum AI.

Abstract

Inspired by the power of large language models (LLMs), our research adapts them to quantum federated learning (QFL) to boost efficiency and performance. We propose a federated fine-tuning method that distills an LLM within QFL, allowing each client to locally adapt the model to its own data while preserving privacy and reducing unnecessary global updates. The fine-tuned LLM also acts as a reinforcement agent, optimizing QFL by adjusting optimizer steps, cutting down communication rounds, and intelligently selecting clients. Experiments show significant efficiency gains. We pioneer a synergy between LLM and QFL, offering: i) practical efficiency: Reduced communication costs and faster convergence. ii) theoretical rigor: Provable guarantees for adaptive federated optimization. iii) scalability: PEFT methods (LoRA, QLoRA) enable deployment on resource-constrained quantum devices. Code implementation is available here 1.

Figures (27)

• Figure 1: Distilling LLMs over QFL: Locally Fine-Tuned LLMs for enhanced QFL Performance
• Figure 2: Execution workflow of LLM-QFL on a real IBM quantum computer, detailing data encoding, quantum circuit selection, and result interpretation.
• Figure 3: Proposed LLM-QFL Framework. Each device fine-tunes its local LLM on its dataset during the initial communication round, followed by training a QCNN on local data. In subsequent rounds, local LLM fine-tuning is skipped, but knowledge distillation from the fine-tuned LLM enhances QCNN adaptation. This enables the QCNN to refine its local optimizer dynamically, leveraging comparative performance analysis to improve efficiency and model convergence.
• Figure 4: Device 0 observations; Decreasing ratio indicates convergence and less gap between performance of LLM model and Quantum Model.
• Figure 5: a) Impact on Device performance. b) Impact on Server performance.

Theorems & Definitions (9)

• Remark 2.1: Lower Loss & Efficient Convergence
• Theorem 6.4: Convergence of LLM-QFL
• proof
• Theorem 6.5: Communication Complexity
• proof
• Theorem 6.6: Computation Complexity
• Remark 6.7
• Corollary 6.8: Efficiency Gains of LLM-QFL
• proof