Pith. sign in

REVIEW 4 major objections 5 minor 56 references

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

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read The paper claims that distilling a locally fine-tuned large language model into quantum federated learning—using the LLM as both teacher and controller—yields an O(1/T) convergence rate while cutting communication rounds and gradient…

desk verdict The LLM-as-controller idea is novel and worth a look, but the convergence proof is a restatement of FedAvg with no handling of the adaptive steps or biased client selection, and the experiments are too weak to support the claims. read the letter →

arxiv 2505.18656 v1 pith:RKUOSF5B submitted 2025-05-24 cs.LG

classification cs.LG
keywords QuantumFederatedLearningLargeLanguageModelsKnowledgeDistillationFine-TuningLoRAQLoAdaptiveOptimizerClientSelection
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Quantum federated learning (QFL) is limited by expensive communication between quantum clients and a server. This paper proposes to give each client a locally fine-tuned large language model (LLM) and to use that LLM twice: as a distillation teacher that keeps local quantum models aligned with the global model, and as a controller that raises the optimizer's iteration budget when a client's loss lags behind the LLM's loss, selects only clients closest to the global loss for aggregation, and stops training when relative improvement is tiny. Under L-smoothness, bounded variance, and strong convexity, the paper claims convergence at rate $O(1/T)$, with communication rounds reduced by $\mathbb{E}[K^t]/K$ and gradient variance by $(1 - k/N)$. Experiments on genomic and tweet-sentiment data, run on simulators and a real quantum processor, show the regulated version converging sooner and in fewer rounds. If correct, the scheme would make quantum federated learning more practical on noisy, resource-limited devices without centralizing private data.

What carries the argument

The load-bearing object is the locally fine-tuned LLM used as both a distillation teacher and a controller. The local update combines gradient descent on the client's quantum model with a distillation term $\lambda K(\theta_g, \theta_i^{t+1})$ that pulls the local model toward the global teacher. The controller uses the loss ratio $L_i^t / L_{\text{LLM}}^t$ to set the number of local optimizer steps, $K_i^t = \text{iter} \cdot L_i^t / L_{\text{LLM}}^t$, so a client whose quantum loss exceeds the LLM benchmark runs more COBYLA iterations; the selection rule is $S_t = \{i : d_i^t \text{ among the smallest } k\%\}$ with $d_i^t = |L_i^t - L_g^t|$. The proof's mechanism is the standard FedAvg error decomposition of equation (9), $\mathbb{E}\|(\theta^{t+1}, \phi^{t+1}) - (\theta^*, \phi^*)\|^2 \leq (1 - \eta_t \mu)\mathbb{E}\|(\theta^t, \phi^t) - (\theta^*, \phi^*)\|^2 + \eta_t^2 (B + C)$, where $B$ regroups client variance, the non-IID gap, and local-step drift and $C$ captures the variance of the selected subset; the paper assumes this decomposition carries over from fixed-step FedAvg to the adaptive-step, alignment-selected setting.

What would settle it

Derive the drift and variance terms in inequality (9) for $K_i^t = \text{iter} \cdot L_i^t / L_{\text{LLM}}^t$ and for $S_t$ selected by smallest $|L_i^t - L_g^t|$; if the combined term $B+C$ grows with the spread of local losses or with the selection threshold, the claimed $O(1/T)$ rate does not follow. As an empirical check, run LLM-QFL and standard QFL with the same total optimizer iterations and compare communication rounds to a fixed accuracy: the adaptive method must need no more rounds.

Watch

Extended reading notes

Core claim

The central claim is that LLM-guided regulation of quantum federated learning improves both convergence speed and communication efficiency. Concretely, each client first fine-tunes a small LLM on its private data with LoRA or QLoRA, and in later rounds the LLM acts as a benchmark: the optimizer's maximum iteration count is scaled by the loss ratio $L_i^t / L_{\text{LLM}}^t$, so lagging clients take more local steps; the server averages only the $k\%$ of clients whose losses are closest to the server loss; and rounds stop when the relative server-loss improvement falls below $\epsilon$. Theorem VI.4 asserts that under L-smoothness, bounded local variance, and $\mu$-strong convexity, the global objective satisfies $\mathbb{E}[F(\theta^T, \phi^T)] - F^* \leq \frac{2L}{\mu}\frac{\Psi}{T + \gamma}$, and Corollary VI.8 translates this into a reduction in communication rounds by $\mathbb{E}[K^t]/K$ and in global-update variance by $(1 - k/N)$ compared with standard QFL. The experiments on genomic and TweetEval sentiment classification, run on simulator backends and on a real IBM quantum processor, are offered as evidence that these mechanisms converge in fewer communication rounds and reduce idle computation.

Load-bearing premise

The argument rests on a standard convergence bound for federated averaging still holding after the paper lets each client run a different number of local steps and averages only the best-aligned clients, without re-deriving the error terms for those changes.

Editorial extensions

If this is right

  • Under the paper's assumptions, the global model reaches $\epsilon$-accuracy in $O\!\left(\frac{L}{\mu}\log\frac{1}{\epsilon} + \frac{B+C}{\mu\epsilon}\right)$ communication rounds, preserving the FedAvg-style rate while adding adaptivity.
  • Adaptive local steps reduce the number of communication rounds by the factor $\mathbb{E}[K^t]/K$ compared with fixed-step QFL, because lagging clients take more optimizer iterations per round.
  • Selecting only the $k\%$ of clients closest to the server loss reduces the variance of the global update by at most $1 - k/N$ relative to random selection, which should make convergence less sensitive to outlier clients.
  • Early stopping when the relative server-loss improvement drops below $\epsilon$ removes rounds that contribute little, lowering total communication cost.
  • LoRA and QLoRA fine-tuning keep the LLM controller small enough to run on quantum-classical devices with limited memory, making the scheme scalable to resource-constrained settings.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The controller pattern—loss-ratio-scaled local steps plus alignment-based client selection—does not depend on the local model being quantum; the same design could be tested in classical federated LLM fine-tuning, where communication cost is also dominated by rounds and uploads.
  • The claimed variance reduction assumes the server can rank clients by a loss that reflects true generalization; a testable extension would compare selection by training loss, validation loss, and LLM-consistency scores to see which proxy preserves the $(1 - k/N)$ benefit.
  • Because the real-hardware experiments were limited to a few communication rounds by usage quotas, a natural follow-up is to run the adaptive controller for many rounds under a fixed noise model; this would show whether early-round gains persist after quantum noise accumulates.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes LLM-QFL, a framework that integrates large language models into quantum federated learning (QFL) by locally fine-tuning an LLM on each client, distilling it into quantum models, and using it as a controller to adjust optimizer iterations, select clients by alignment with the global model, and stop training early. The authors claim a convergence rate of O(1/T) under L-smoothness, bounded variance, and strong convexity, as well as communication and variance reductions. Experiments are reported on genomic (DemoHumanOrWorm) and TweetEval-sentiment datasets using VQC and QCNN models, with simulations and real IBM quantum hardware.

Significance. If the theoretical claims were valid, the paper would offer a useful bridge between LLM-based adaptive control and quantum federated learning, with practical implications for communication efficiency and client selection. The authors deserve credit for running experiments on real IBM quantum hardware and for releasing code, which is uncommon and valuable. However, the central theoretical contribution is not established: the convergence theorem is imported from standard FedAvg analysis without addressing the adaptive local steps and biased client selection that define the algorithm, and a key remark assumes the very inequality the paper aims to prove, while an appendix figure contradicts it. The experimental evidence, while suggestive, is also limited by near-chance accuracies on the genomic task and a small number of communication rounds on real hardware.

major comments (4)
  1. [Appendix A, Theorem VI.4] The proof of Theorem VI.4 transfers the FedAvg convergence bound of Li et al. [45] to Algorithm 1, but the transfer is not justified: equation (8) defines local updates with an adaptive number of local steps K_i^t = iter * L_i^t / L_LLM and aggregation over the alignment-selected set S_t = argmin_{|S|=K} sum_{i in S} |L_i^t - L_g^t|, whereas the bound in [45] requires a fixed number E of local steps and uniform random (or full) participation to keep the aggregated update unbiased and to control client drift. Equation (9) is asserted without deriving the drift and variance terms for these modified choices, so the claimed O(1/T) rate is unsupported.
  2. [Appendix A, Theorem VI.4, constants B and C] The constants B and C in Theorem VI.4 contain terms (E-1)^2 G^2 and E^2 G^2, but the adaptive step count K_i^t = iter * L_i^t / L_LLM is not bounded above unless L_LLM is uniformly bounded away from zero; no such control is stated or proved. Consequently, inequality (9) cannot be obtained from the fixed-E derivation in [45], and the theorem's statement is not supported by the proof.
  3. [Section II.B, Remark II.1] Remark II.1 assumes the inequality l(·,·) < l_LLM(·,·) as a consequence of distillation, but this inequality is the central performance claim the paper aims to establish, so the remark is circular. Moreover, Figure 23 in Appendix H reports that the LLM's loss is lower than the quantum model's loss in all three settings, which directly contradicts the assumed inequality and further undermines the remark's validity.
  4. [Appendix A, Corollary VI.8] The efficiency corollary is not established. Claim 1 reduces the communication-round improvement to the definitional identity E[K_i^t] = E[iter * L_i^t / L_LLM] >= K, which holds only when L_i^t >= L_LLM; it does not follow from Theorem VI.4 because the theorem did not prove a convergence rate for the adaptive-selection algorithm. Claim 2 bounds Var(∇F_LLM-QFL) by (1 - k/N) Var(∇F_QFL) using a Markov-type argument on sorted deviations d_i^t, but the variance term in the FedAvg bound is not this population quantity, and the selection rule makes the aggregated gradient biased, so the comparison with random selection is not valid.
minor comments (5)
  1. [Algorithm 1, line 22] The notation N' = n/N is unclear; this appears intended to denote a count of selected devices, but as written it is a fraction of N, and the subsequent averaging over N' devices is ambiguous.
  2. [Equation (3)] Equation (3) uses the same index i on both sides of the update, so the expression theta^{t+1}_i = sum_i w_i theta^t_i is not well defined; the aggregation should use distinct client indices.
  3. [Table I] The reported train and test accuracies for the genomic task are close to 0.5 for several simulator and real-hardware runs, which is near chance level; the authors should discuss whether the method actually learns the task under those settings.
  4. [Section IV.C and Appendix C] There is an inconsistency in the spelling of the IBM hardware name ('IBM Brisbane' in the main text and 'IBM Bribane' in Appendix C), and the paper should use a consistent citation for the LLaMA model, since reference [35] is the original LLaMA paper rather than the Meta-LLAMA 3.2-1B model described in the experiments.
  5. [Figures 4–10] Several convergence and performance figures (e.g., Figures 4, 5, 8, 9) report single runs without error bars or statistical significance, which makes it difficult to assess the robustness of the claimed efficiency gains.

Circularity Check

3 steps flagged · score 6.0 of 10

The central convergence theorem imports the FedAvg bound with renamed adaptive-step and client-selection terms, and the efficiency corollaries read the claimed gains off their own definitions.

  1. self definitional [Section II.B, Remark II.1]
    "By acting as a knowledge distiller, the LLM guides QFL training, improving convergence and efficiency while reducing computational overhead. Therefore, we have an inequality l(·,·)< l LLM(·,·) which implies that the QFL-trained quantum model achieves a lower loss than an LLM used in isolation."

    The inequality l < l_LLM is the central benefit the paper claims to establish, but it is asserted as an immediate consequence of the phrase 'acting as a knowledge distiller' rather than derived from the algorithm, the convergence theorem, or the measurements. The paper's own Figure 23 states that 'the performance of LLM is unmatched,' i.e., l_LLM < l in the experiments, so the assumed inequality is exactly the contested conclusion, not a premise supplied by the method.

  2. renaming known result [Appendix A, Theorem VI.4, proof equations (8)-(9)]
    "With error decomposed, we can have E∥(θt+1, ϕt+1)−(θ ∗, ϕ∗)∥2 ≤(1−η tµ)E∥ (θt, ϕt)−(θ ∗, ϕ∗)∥2 +η 2 t (B+C)(9) ... B= PM i=1 w2 i σ2 i + 6LΓ + 8(E−1) 2G2, and C= 4 St E2G2 ... Et i =iter∗ Lt i LLLM for i∈S t and St = arg min S⊆{1,...,N},|S|=K P i∈S |Lt i − Lt g|."

    Equation (9) is the FedAvg recursion of Li et al. [45] with the fixed local-step count E replaced by K_i^t = iter * L_i^t / L_LLM and the client set replaced by S_t = argmin sum |L_i^t - L_g^t|. The proof states the decomposition without re-deriving the drift and variance terms B and C for loss-dependent step counts or for selection that is correlated with current client losses; such selection biases the aggregated gradient estimator, and the (E-1)^2 and E^2 terms do not follow. The O(1/T) rate is therefore the imported FedAvg result with renamed symbols, not a bound proved for Algorithm 1.

1 more flagged steps
  1. self definitional [Appendix A, Corollary VI.8.1]
    "TQFL TLLM-QFL ≥ E[K t i ] K ... The adaptive mechanism ensures: E[K (t) i ] =E iter· Lt i Lt LLM ≥K whenever L t i ≥L t LLM . Thus, TLLM-QFL =O 1 E[K t i ] ≤ K E[K t i ] TQFL"

    The claimed communication-round reduction is read directly off the definition K_i^t = iter * L_i^t / L_LLM. The proof assumes the round count scales as the reciprocal of the average number of local steps, but the very bound imported in Theorem VI.4 contains 8(E-1)^2 G^2 and E^2 G^2 drift terms, which grow with E; more adaptive local steps can increase drift and slow convergence in the non-iid setting. The ratio E[K]/K is the designed input of the algorithm, not a derived or measured efficiency gain.

full rationale

The paper's theoretical section is largely a relabeling exercise. Theorem VI.4 cites [45] and writes the FedAvg error recursion as equation (9) after replacing E with the adaptive K_i^t and the participation rule with alignment-based S_t; no proof shows that the FedAvg drift and variance bounds survive these replacements. Corollary VI.8.1 then converts the definition of K_i^t into a claimed communication-round reduction, and Corollary VI.8.2 asserts a (1 - k/N) variance reduction via a Markov step that is not justified and is false in general when the d_i^2 are equal. Separately, Remark II.1 assumes the inequality l < l_LLM that the framework is supposed to deliver, and Figure 23 contradicts it. These are not harmless self-citations: the load-bearing theoretical claims reduce to an imported theorem with renamed terms or to definitions of the adaptive mechanism. The paper does include genuine experiments on simulators and IBM hardware with external datasets and baselines, so the experimental portion is not circular and can stand on its own; the external FedAvg bound is also real evidence, though not for this modified algorithm. The circularity score is therefore intermediate: partial, because the empirical comparisons and the external theorem are real inputs, but the central convergence and efficiency claims as stated are not established by the derivation chain.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The framework introduces no new physical entities. The load-bearing assumptions are standard convex optimization conditions that are questionable for quantum models, plus an ad hoc inequality that is contradicted by the paper's own data. The proof transfers an external FedAvg result to a modified algorithm without justification. The free parameters are hyperparameters that are either unspecified or hand-chosen, and the central variance-reduction claim depends on the unspecified selection fraction.

free parameters (5)
  • lambda (distillation weight) = not specified
    Introduced in Eq. (5) and Eq. (6) to balance distillation loss, but no value is given and it is not used in Algorithm 1 steps.
  • mu (regularization weight) = not specified
    Introduced in Eq. (6) for the regularization term; no value is provided and the term does not appear in the algorithm description.
  • epsilon (termination threshold) = not specified
    Used in termination criterion Delta L_s / L_s < epsilon; the threshold value is not reported.
  • client selection fraction k/N = 10% in experiments
    The selected client fraction is a chosen hyperparameter, claimed to reduce variance by (1 - k/N), but no sensitivity analysis is given.
  • initial optimizer maxiter = 10 (Exp. I) and 5 (Exp. 2 on real hardware)
    The initial iteration limit for COBYLA is chosen by hand; the adaptive mechanism then scales it based on the loss ratio.
assumptions (5)
  • standard math Assumption VI.1: Each local loss function is L-smooth.
    Standard smoothness condition used in optimization theory; plausible for smooth loss surfaces but not verified for QNNs.
  • standard math Assumption VI.2: Bounded gradient variance and bounded expected squared norm of gradients.
    Standard bounded variance assumptions in stochastic optimization; required for the cited FedAvg proof.
  • domain assumption Assumption VI.3: The global loss function is mu-strongly convex.
    Strong convexity is unlikely to hold for variational quantum circuits and QCNNs, which are non-convex in practice. This is a strong domain assumption not justified in the paper.
  • ad hoc to paper Remark II.1: The optimized LLM-QFL loss satisfies l(.,.) < l_LLM(.,.).
    This inequality is stated as a consequence of distillation, but it is the target result. The paper's own Figure 23 shows the LLM outperforming the quantum model, contradicting this assertion.
  • ad hoc to paper The FedAvg convergence bound of Li et al. [45] remains valid when local steps are adaptive and clients are selected by alignment.
    The proof of Theorem VI.4 jumps from equation (8) to the error decomposition in equation (9) without re-deriving the bound for these modified algorithmic choices. This transferability is asserted, not proven.

how reviews work

0 comments
Cite this review

Pith. "Pith review of LLM-QFL: Distilling Large Language Model for Quantum Federated Learning." pith.science (2026). https://pith.science/paper/RKUOSF5B

@misc{pith2026250518656,
  author       = {Pith},
  title        = {Pith review of: LLM-QFL: Distilling Large Language Model for Quantum Federated Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RKUOSF5B}},
  note         = {Machine review of arXiv:2505.18656}
}
read the original 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

Figures reproduced from arXiv: 2505.18656 by the authors.

Figure 1
Figure 1. Distilling LLMs over QFL: Locally Fine￾Tuned LLMs for enhanced QFL Performance Despite their potential, key challenges remain in integrating LLMs within the QFL framework. The feasibility of leveraging classically trained LLMs to enhance quantum machine learning remains unexplored, and optimizing QFL algorithms through LLM fine￾tuning to mitigate computational and communication bottlenecks requires further investiga… view at source ↗
Figure 2
Figure 2. Execution workflow of LLM-QFL on a real IBM quantum computer, detailing data encoding, quantum circuit selection, and result interpretation. to assess the practical feasibility of LLM-QFL while addressing the real-world challenges associated with quantum computation. By integrating LLMs within QFL, we bridge the gap between LLM, quantum intelligence and federated learning, paving the way for a scalable, efficient, a… view at source ↗
Figure 3
Figure 3. Proposed LLM-QFL Framework. Each device fine-tunes its local LLM on its dataset during the [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (18 more)
Figure 4
Figure 4. Figure 4: Device 0 observations; Decreasing ratio indi [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 7
Figure 7. Figure 7: Device 8 performance: All vs Selected (a) Test Acc (b) Val Acc (c) Val Loss [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: Server performance: QFL, LLM-QFL-all vs. [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]
Figure 9
Figure 9. Figure 9: Server Performance: Simulators vs Real Quantum Computer The results of performance in real quantum computer on device accuracy, test and loss in shown in Figures 10a, 10b and 10c. In all instances, the performance starts lower than those in simulators. This highlights …
Figure 10
Figure 10. Figure 10: Device Performance: Simulators vs Real Quantum Computer Due to limited usage limit, the results on real quan￾tum computer are only for two (Exp 1) communication rounds. The number of devices used is 4 including server device. The maximum maxiter value is 5 with the in…
Figure 11
Figure 11. Figure 11: Usage times and Timestamps on execution on IBM quantum Computer. We further performed experimental analysis (Exp 2) with following settings. Dataset used: Genomic; both train and test set downloaded from the DemoHuman￾Worm Dataset. Train size used is 1000 and test siz…
Figure 12
Figure 12. Figure 12: TweetEval-Sentiment Train Set 3) Data Preparation of LLM fine-tuning and VQC training: Both the architecture of LLM and VQC is different. LLM is based on the Transformer architecture, while VQC is based on the Quantum Neural Network architecture. We cannot feed the sa…
Figure 13
Figure 13. Figure 13: Encoded and PCA applied Genomic Dataset IBMRuntimeError: ’Failed to run program: \’403 Client Error: Forbidden for url: https://api.quantum.ibm.com/runtime/jobs. {"errors":[{"message":"Job create exceeds open plan job usage limits","code ":4317,"solution":"Please wait…
Figure 14
Figure 14. Figure 14: Quantum Convolutional Neural Network (QCNN); Example of QCNN with 4 qubits. At each layer, [PITH_FULL_IMAGE:figures/full_fig_p014_14.png]
Figure 15
Figure 15. Figure 15: Variational Quantum Classifier with 4 Qubits; Consists of feature map using ZZFeatureMap, ansatz [PITH_FULL_IMAGE:figures/full_fig_p014_15.png]
Figure 18
Figure 18. Figure 18: Measurements Counts [PITH_FULL_IMAGE:figures/full_fig_p014_18.png]
Figure 19
Figure 19. Figure 19: Convolutional and pooling circuit. of feature map for data encoding, ansatz with trainable parameters etc. as shown in [PITH_FULL_IMAGE:figures/full_fig_p014_19.png]
Figure 20
Figure 20. Figure 20: Impact of choice of regularization adjustment, logarithmic, adaptive, etc. With an incre￾mental approach, we compute the ratio between the current performance of the device and the loss value of the LLM evaluation, calculate the increment to align performance, and inc…
Figure 23
Figure 23. Figure 23: Device performance comparison of results [PITH_FULL_IMAGE:figures/full_fig_p015_23.png]
Figure 24
Figure 24. Figure 24: Device loss and LLM f1 Score. I. Other Results [PITH_FULL_IMAGE:figures/full_fig_p015_24.png]
Figure 22
Figure 22. Figure 22: Average Device performance; QFL, LLM￾QFL-all vs. LLM-QFL-selected. H. LLM vs QFL comparison In this work, we fine-tune LLM models and use their performance to regulate and further manage the QFL algorithm autonomously in the direction and with the intention of improvi…
Figure 25
Figure 25. Figure 25: Average Device performance: QFL vs. LLM [PITH_FULL_IMAGE:figures/full_fig_p016_25.png]
Figure 27
Figure 27. Figure 27: shows results in terms of the cumulative time taken by jobs to start and finish and their trend throughout job executions on the IBM computer. This shows that there are subtle variations in the way jobs are executed on the IBM computer [PITH_FULL_IMAGE:figures/full_f…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

56 extracted references · 44 canonical work pages

  1. [45]

    ON THE CONVERGENCE OF FEDA VG ON NON-IID,

    X. Li, W. Yang, K. Huang, S. Wang, and Z. Zhang, “ON THE CONVERGENCE OF FEDA VG ON NON-IID,”ICLR, 2020. JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2025 11 APPENDIX A. CONVERGENCE& COMPLEXITY OFLLM-QFL We made the following assumptions. Assumption VI.1(Lipschitz Smoothness).Each local loss functionF i(θ, ϕ)is L-smooth: ∥∇Fi(θ, ϕ)− ∇Fi(θ′, ϕ′)∥ ≤L...

  2. [1]

    OpenFedLLM: Training Large Language Models on Decentralized Private Data via Federated Learning,

    R. Yeet al., “OpenFedLLM: Training Large Language Models on Decentralized Private Data via Federated Learning,” Feb. 2024, arXiv:2402.06954 [cs]

  3. [2]

    Towards Quantum Machine Learn- ing for Constrained Combinatorial Optimization: a Quantum QAP Solver,

    X. Ye, G. Yan, and J. Yan, “Towards Quantum Machine Learn- ing for Constrained Combinatorial Optimization: a Quantum QAP Solver,”ICML, 2023

  4. [3]

    Quantum Support Vector Machine for Classifying Noisy Data,

    J. Li, Y . Li, J. Song, J. Zhang, and S. Zhang, “Quantum Support Vector Machine for Classifying Noisy Data,”IEEE Transactions on Computers, vol. 73, no. 9, pp. 2233–2247, Sep. 2024

  5. [4]

    Quantum federated learning experiments in the cloud with data encoding,

    S. R. Pokhrel, N. Yash, J. Kua, G. Li, and L. Pan, “Quantum federated learning experiments in the cloud with data encoding,” arXiv preprint arXiv:2405.00909, 2024

  6. [5]

    A Survey on Multimodal Large Language Models,

    S. Yin, C. Fu, S. Zhao, K. Li, X. Sun, T. Xu, and E. Chen, “A Survey on Multimodal Large Language Models,” Apr. 2024, arXiv:2306.13549 [cs]

  7. [6]

    Deepseek-r1: Incentivizing reasoning capa- bility in llms via reinforcement learning,

    D. Guoet al., “Deepseek-r1: Incentivizing reasoning capa- bility in llms via reinforcement learning,”arXiv preprint arXiv:2501.12948, 2025

  8. [7]

    NetLLM: Adapting Large Language Models for Networking

    D. Wu, X. Wang, Y . Qiao, Z. Wang, J. Jiang, S. Cui, and F. Wang, “NetLLM: Adapting Large Language Models for Networking,” inProceedings of the ACM SIGCOMM 2024 Conference, Aug. 2024, pp. 661–678, arXiv:2402.02338 [cs]

Show all 56 references
  1. [8]

    FedBiOT: LLM Local Fine-tuning in Federated Learning without Full Model,

    F. Wu, Z. Li, Y . Li, B. Ding, and J. Gao, “FedBiOT: LLM Local Fine-tuning in Federated Learning without Full Model,” inProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, ser. KDD ’24. New York, NY , USA: Association for Computing Machinery, ...

  2. [9]

    A data- encoding approach to quantum federated learning: Experiment- ing with cloud challenges,

    S. R. Pokhrel, N. Yash, J. Kua, G. Li, and L. Pan, “A data- encoding approach to quantum federated learning: Experiment- ing with cloud challenges,” inProceedings of the 8th Asia- Pacific Workshop on Networking, 2024, pp. 179–180

  3. [10]

    QuantumFed: A Federated Learning Framework for Collaborative Quantum Training,

    Q. Xia and Q. Li, “QuantumFed: A Federated Learning Framework for Collaborative Quantum Training,” in2021 IEEE Global Communications Conference (GLOBECOM), Dec. 2021, pp. 1–6

  4. [11]

    Quantum Federated Learning with Quantum Data,

    M. Chehimi and W. Saad, “Quantum Federated Learning with Quantum Data,” inICASSP 2022 - 2022 IEEE Interna- tional Conference on Acoustics, Speech and Signal Processing (ICASSP), May 2022, pp. 8617–8621

  5. [12]

    Quantum federated learning through blind quantum computing,

    W. Li, S. Lu, and D.-L. Deng, “Quantum federated learning through blind quantum computing,”Science China Physics, Mechanics & Astronomy, vol. 64, no. 10, p. 100312, Oct. 2021

  6. [13]

    Federated quantum natural gradient descent for quantum federated learning,

    J. Qi and M.-H. Hsieh, “Federated quantum natural gradient descent for quantum federated learning,” pp. 329–341, 2024

  7. [14]

    Slimmable quantum federated learning,

    W. J. Yun, J. P. Kim, S. Jung, J. Park, M. Bennis, and J. Kim, “Slimmable quantum federated learning,”arXiv preprint arXiv:2207.10221, 2022

  8. [15]

    Federated Quantum Machine Learning,

    S. Y .-C. Chen and S. Yoo, “Federated Quantum Machine Learning,”Entropy, vol. 23, no. 4, p. 460, Apr. 2021

  9. [16]

    Exact Decomposition of Quantum Channels for Non-IID Quantum Federated Learning,

    H. Zhao, “Exact Decomposition of Quantum Channels for Non-IID Quantum Federated Learning,” Sep. 2022, arXiv:2209.00768

  10. [17]

    Quantum Federated Learning With Decentralized Data,

    R. Huang, X. Tan, and Q. Xu, “Quantum Federated Learning With Decentralized Data,”IEEE Journal of Selected Topics in Quantum Electronics, vol. 28, no. 4: Mach. Learn. in Photon. Commun. and Meas. Syst., pp. 1–10, Jul. 2022

  11. [18]

    A personalized quantum federated learning,

    D. Gurung and S. R. Pokhrel, “A personalized quantum federated learning,” inProceedings of the 8th Asia-Pacific Workshop on Networking, 2024, pp. 175–176

  12. [19]

    Chained continuous quantum federated learning frame- work,

    ——, “Chained continuous quantum federated learning frame- work,”Future Generation Computer Systems, p. 107800, Mar. 2025

  13. [20]

    Quantum federated learning for metaverse: Analysis, design and implementation,

    D. Gurung, S. R. Pokhrel, and G. Li, “Quantum federated learning for metaverse: Analysis, design and implementation,” IEEE Transactions on Network and Service Management, pp. 1–1, 2025

  14. [21]

    Performance Analysis and De- sign of a Weighted Personalized Quantum Federated Learning,

    D. Gurung and S. R. Pokhrel, “Performance Analysis and De- sign of a Weighted Personalized Quantum Federated Learning,” IEEE Transactions on Artificial Intelligence, pp. 1–12, 2025

  15. [22]

    PrE-Text: Training Language Models on Private Federated Data in the Age of LLMs,

    C. Hou, A. Shrivastava, H. Zhan, R. Conway, T. Le, A. Sagar, G. Fanti, and D. Lazar, “PrE-Text: Training Language Models on Private Federated Data in the Age of LLMs,” inProceedings of the 41st International Conference on Machine Learning. PMLR, Jul. 2024, pp. 19 043–19 061

  16. [23]

    FedBPT: Efficient Federated Black-box Prompt Tuning for Large Language Models,

    J. Sun, Z. Xu, H. Yin, D. Yang, D. Xu, Y . Liu, Z. Du, Y . Chen, and H. R. Roth, “FedBPT: Efficient Federated Black-box Prompt Tuning for Large Language Models,” inProceedings of the 41st International Conference on Machine Learning. PMLR, Jul. 2024, pp. 47 159–47 173

  17. [24]

    Federated Full-Parameter Tuning of Billion-Sized Language Models with Communication Cost under 18 Kilobytes,

    Z. Qin, D. Chen, B. Qian, B. Ding, Y . Li, and S. Deng, “Federated Full-Parameter Tuning of Billion-Sized Language Models with Communication Cost under 18 Kilobytes,” in Proceedings of the 41st International Conference on Machine Learning. PMLR, Jul. 2024, pp. 41 473–41 497

  18. [25]

    FDLoRA: Personalized Federated Learning of Large Language Model via Dual LoRA Tuning,

    J. QI, Z. Luan, S. Huang, C. Fung, H. Yang, and D. Qian, “FDLoRA: Personalized Federated Learning of Large Language Model via Dual LoRA Tuning,” Jun. 2024, arXiv:2406.07925 [cs]

  19. [26]

    FedDAT: An Approach for Foundation Model Finetuning in Multi-Modal Heterogeneous Federated Learning,

    H. Chen, Y . Zhang, D. Krompass, J. Gu, and V . Tresp, “FedDAT: An Approach for Foundation Model Finetuning in Multi-Modal Heterogeneous Federated Learning,”Proceedings of the AAAI Conference on Artificial Intelligence, vol. 38, no. 10, pp. 11 285– 11 293, Mar. 2024, number: 10

  20. [27]

    FLoRA: Federated Fine-Tuning Large Language Models with Heterogeneous Low-Rank Adaptations,

    Z. Wang, Z. Shen, Y . He, G. Sun, H. Wang, L. Lyu, and A. Li, “FLoRA: Federated Fine-Tuning Large Language Models with Heterogeneous Low-Rank Adaptations,” Sep. 2024, arXiv:2409.05976 [cs]

  21. [28]

    FedPETuning: When federated learning meets the parameter- efficient tuning methods of pre-trained language models,

    Z. Zhang, Y . Yang, Y . Dai, Q. Wang, Y . Yu, L. Qu, and Z. Xu, “FedPETuning: When federated learning meets the parameter- efficient tuning methods of pre-trained language models,” in Findings of the Association for Computational Linguistics: ACL

  22. [29]

    pFedPrompt: Learning Per- sonalized Prompt for Vision-Language Models in Federated Learning,

    T. Guo, S. Guo, and J. Wang, “pFedPrompt: Learning Per- sonalized Prompt for Vision-Language Models in Federated Learning,” inProceedings of the ACM Web Conference 2023, ser. WWW ’23. New York, NY , USA: Association for Computing Machinery, Apr. 2023, pp. 1364–1374

  23. [30]

    Scaling Federated Learning for Fine-Tuning of Large Language Models,

    A. Hilmkil, S. Callh, M. Barbieri, L. R. S ¨utfeld, E. L. Zec, and O. Mogren, “Scaling Federated Learning for Fine-Tuning of Large Language Models,” inNatural Language Processing and Information Systems, E. M ´etais, F. Meziane, H. Horacek, and E. Kapetanios, Eds. Cham: Spring...

  24. [31]

    Titanic: Towards Production Federated Learning with Large Language Models,

    N. Su, C. Hu, B. Li, and B. Li, “Titanic: Towards Production Federated Learning with Large Language Models,” inIEEE INFOCOM 2024 - IEEE Conference on Computer Commu- nications. Vancouver, BC, Canada: IEEE, May 2024, pp. 611–620

  25. [32]

    Recovering Private Text in Federated Learning of Language Models,

    S. Gupta, Y . Huang, Z. Zhong, T. Gao, K. Li, and D. Chen, “Recovering Private Text in Federated Learning of Language Models,”Advances in Neural Information Processing Systems, vol. 35, pp. 8130–8143, Dec. 2022

  26. [33]

    Genomic benchmarks: a collection of datasets for genomic sequence classification,

    K. Greˇsov´a, V . Martinek, D.ˇCech´ak, P. ˇSimeˇcek, and P. Alexiou, “Genomic benchmarks: a collection of datasets for genomic sequence classification,”BMC Genomic Data, vol. 24, no. 1, p. 25, 2023

  27. [34]

    Semeval-2017 task 4: Sentiment analysis in twitter,

    S. Rosenthal, N. Farra, and P. Nakov, “Semeval-2017 task 4: Sentiment analysis in twitter,” inProceedings of the 11th international workshop on semantic evaluation (SemEval-2017), 2017, pp. 502–518

  28. [35]

    Llama: Open and efficient foundation language models,

    H. Touvronet al., “Llama: Open and efficient foundation language models,” 2023

  29. [36]

    Language models are unsupervised multitask learners,

    A. Radford, J. Wu, R. Child, D. Luan, D. Amodei, and I. Sutskever, “Language models are unsupervised multitask learners,” 2019

  30. [37]

    Attention is All you Need,

    A. Vaswaniet al., “Attention is All you Need,” inAdvances in Neural Information Processing Systems, vol. 30. Curran Associates, Inc., 2017

  31. [38]

    A Survey of Large Language Models,

    W. X. Zhaoet al., “A Survey of Large Language Models,” Nov. 2023, arXiv:2303.18223 [cs]

  32. [39]

    Parameter- Efficient Fine-Tuning for Large Models: A Comprehensive Survey,

    Z. Han, C. Gao, J. Liu, J. Zhang, and S. Q. Zhang, “Parameter- Efficient Fine-Tuning for Large Models: A Comprehensive Survey,” Sep. 2024, arXiv:2403.14608

  33. [40]

    LoRA: Low-Rank Adaptation of Large Language Models,

    E. J. Hu, Y . Shen, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, and W. Chen, “LoRA: Low-Rank Adaptation of Large Language Models,” Oct. 2021, arXiv:2106.09685

  34. [41]

    Distilling the Knowledge in a Neural Network,

    G. Hinton, O. Vinyals, and J. Dean, “Distilling the Knowledge in a Neural Network,” Mar. 2015

  35. [42]

    Model compression,

    C. Bucilu ˇa, R. Caruana, and A. Niculescu-Mizil, “Model compression,” inProceedings of the 12th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, ser. KDD ’06. New York, NY , USA: Association for Computing Machinery, Aug. 2006, pp. 535–541

  36. [43]

    A Practical Recipe for Federated Learning Under Statistical Heterogeneity Experimen- tal Design,

    M. Morafah, W. Wang, and B. Lin, “A Practical Recipe for Federated Learning Under Statistical Heterogeneity Experimen- tal Design,”IEEE Transactions on Artificial Intelligence, vol. 5, no. 4, pp. 1708–1717, Apr. 2024

  37. [44]

    Communication-efficient federated learning via knowledge distillation,

    C. Wuet al., “Communication-efficient federated learning via knowledge distillation,”Nature Communications, vol. 13, no. 1, p. 2032, Apr. 2022

  38. [47]

    (Adaptive Step Size Efficiency) TQFL TLLM-QFL ≥ E[K t i ] K where, K is the fixed number of local steps in QFL and E[K (t) i ]is the average adaptive steps in LLM-QFL

  39. [48]

    (Variance Reduction) Var(∇FLLM-QFL)≤ 1− k N Var(∇FQFL) where,k/Nis the fraction of selected clients. Proof. 1. From the convergence rate in Theorem 1, the effective progress per round scales with the number of local steps. The adaptive mechanism ensures: E[K (t) i ] =E iter· L...

  40. [49]

    For client selection variance, let St rand be random selection, and St align be our alignment-based selection. The variance decomposes as, Var(∇F) =E[∥∇F i − ∇F∥2] = 1 N NX i=1 d(t)2 i Our selection criterion S(t) ={i|d t i smallestk%} minimizes Var(∇FLLM-QFL) = 1 k X i∈S(t) a...

  41. [50]

    For thesentencetask, there are three classesnegative, positive,andneutral with varying sample as shown in Figure 12

    TweetEval, Sentiment:TweetEval consists of het- erogeneous tasks in Twitter which is used as multiclass dataset with tasks likeirony, hate, offensive, stance, emoji, emotionandsentiment. For thesentencetask, there are three classesnegative, positive,andneutral with varying sam...

  42. [51]

    The encoded and PCA (n=4) applied dataset is shown in Figure 13

    Genomic-DemoHumanOrWorm Dataset:Demo- HumanOrWorm is a benchmark genomic sequence dataset for classification purposes which consists of 75,000 train samples and 25,000 test set samples with data instance as, (’TCAACTGACTTCCGAGGGAATAAGTGTTTCGCCAT CTCGAACTGTATACTCTGCTATCAA GACCG...

  43. [52]

    LLM is based on the Transformer architecture, while VQC is based on the Quantum Neural Network architecture

    Data Preparation of LLM fine-tuning and VQC training:Both the architecture of LLM and VQC is different. LLM is based on the Transformer architecture, while VQC is based on the Quantum Neural Network architecture. We cannot feed the same data format to both models. Thus, there ...

  44. [53]

    Based on the required data sample size, we deduct small samples

    First, we load and shuffle dataset. Based on the required data sample size, we deduct small samples

  45. [54]

    dset” for train or “test

    Then, we convert dataset into dataframe with columns “dset” for train or “test” set, cat to store labels (Human or Worm), “seq” for nucleotide sequences

  46. [55]

    For LLaMA model, first we convert dataframe to HuggingFace Dataset format. For tokenization, we first define k-mer tokenization (substrings of length k=6), load LLaMA tokenizer (meta-llama/Llama- 3.2-1B), ensure padding, load pretrained model for sequence classification (2 lab...

  47. [56]

    Job create exceeds open plan job usage limits

    For VQC model, one-hot encoding is applied to nucleotide sequences as A=[1,0,0,0], C=[0,1,0,0], G=[0,0,1,0], T=[0,0,0,1] and are converted to one- hot vectors. Also, the Principal component analysis is applied for dimensionality reduction with PCA (n components=4) which reduce...

  48. [2023]

    Association for Computational Linguistics (ACL), 2023

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.