Abstract
Federated fine-tuning of large language models (LLMs) over bandwidth-limited 6G links must meet strict round-time and energy budgets. Analog over-the-air (OTA) aggregation reduces uplink cost but is sensitive to fading and interference, which distort the aggregated gradient. We consider a two-phase workflow (centralized pre-training followed by federated fine-tuning) where the base station uses a movable-antenna (MA) array. In each round, MA element positions and the receive/transmit beamformers are adjusted under minimum-spacing constraints to reshape the channel and improve OTA aggregation without increasing user transmit power. We formulate a mixed-integer, nonconvex resource-allocation problem that jointly selects clients and optimizes the number of global rounds, CPU frequencies, mini-batch sizes, MA positions, and analog weights under end-to-end latency and energy limits. A successive convex approximation-penalty dual decomposition (SCA-PDD) routine alternates convex updates with oblique-manifold beamforming and spacing-aware MA placement. Experiments on OpenLLaMA-v2 (3B) with LoRA and 4-bit quantization on Alpaca and Dolly (10 clients) attain round-30 validation perplexities as low as 2.94 (Alpaca, K=1) and 4.62 (Dolly, K=1). Relative to the strongest non-MA baseline at the same concurrency, this corresponds to 17.4 percent (Alpaca, K=1) and 54.4 percent (Dolly, K=1) lower perplexity; at K=2 the reductions are 14.2 percent (Alpaca) and 13.7 percent (Dolly). Participation fairness also improves across all uplink concurrencies K in {1,2,4,8}, with the largest margins when fewer clients transmit per round.
Introduction
The deployment of Large Language Models (LLMs) at the network edge is a cornerstone of the 6G vision, enabling low-latency, privacy-preserving AI services. However, fine-tuning these models on edge devices is severely constrained by the limited bandwidth and energy of wireless links. Federated Learning (FL) offers a solution by keeping data local, but the communication overhead of exchanging massive model updates remains a significant bottleneck.
To mitigate this, Over-the-Air (OTA) aggregation has emerged as a promising technique. By allowing devices to transmit simultaneously on the same frequency, the server receives a superposition of signals that naturally computes the weighted sum of gradients. Yet, OTA aggregation is highly susceptible to "analog-sum mismatch" caused by wireless fading and interference. This paper introduces a novel physical-layer solution: Movable Antennas (MA). By physically repositioning antenna elements at the base station, we can "reshape" the wireless channel to perfectly align incoming signals, significantly enhancing the stability and accuracy of federated LLM fine-tuning.
System Architecture
The proposed framework operates in a two-phase workflow designed to bridge the gap between general-purpose pre-training and domain-specific edge applications.
Phase I: Centralized Pre-Training
In the first phase, a base model is pre-trained on a large, general dataset ($D_{gen}$) to minimize the population loss $L_{pre}(\theta)$. This phase establishes the initial weights ($\theta_0$), tokenizer, and vocabulary that serve as the foundation for subsequent fine-tuning. Decisions in this phase include setting the batch size $W^{(m)}$ and server frequency $f^{(m)}$ to manage time and energy budgets ($L_{max}, Q_{max}$).
Phase II: Task-Specific Federated Fine-Tuning
The second phase involves distributed clients (e.g., mobile devices, laptops) fine-tuning the model on domain-specific local data. The base station, equipped with a movable antenna array, manages the global update process. Unlike traditional fixed-antenna systems, the MA array can adjust its geometry in each round to optimize the reception of gradients from the selected set of clients ($U_n$).
Technical Innovations
Movable Antenna (MA) Array
The core innovation lies in treating the physical position of antenna elements as a controllable degree of freedom. By moving antennas within a local region, the system can exploit the spatial variation of the wireless channel. This allows for precise phase and amplitude alignment of the uplink gradients, minimizing the noise variance ($\sigma^2$) in the OTA aggregation process without requiring additional transmit power from energy-constrained edge devices.
Joint Resource Allocation and SCA-PDD Algorithm
The system optimizes a complex set of parameters, including the number of training rounds, client CPU frequencies, mini-batch sizes, and the physical coordinates of the MA elements. This results in a mixed-integer, nonconvex optimization problem. To solve this, the authors developed a hybrid Successive Convex Approximation-Penalty Dual Decomposition (SCA-PDD) algorithm.
- SCA is used to handle continuous variables like batch sizes and beamforming weights.
- PDD manages the discrete constraints associated with client selection and the minimum spacing requirements ($d_{min}$) between movable antenna elements.
Convergence Analysis with Distribution Shift
Recognizing that edge data often differs significantly from pre-training data, the authors incorporated Wasserstein distance into their convergence analysis. This mathematical framework accounts for the "distribution shift," ensuring that the fine-tuning process remains stable even when local datasets are highly heterogeneous.
Experimental Evaluation
The framework was evaluated using the OpenLLaMA-v2 (3B) model. To ensure feasibility on edge hardware, the researchers employed LoRA (Low-Rank Adaptation) and 4-bit quantization. The system was tested on two instruction-tuning datasets: Alpaca (52k pairs) and Dolly (15k pairs), distributed across 10 simulated clients.
Performance and Quality
The results demonstrate a substantial improvement in model quality. At round 30, the MA-enhanced system achieved validation perplexities as low as 2.94 on Alpaca and 4.62 on Dolly. Compared to the best non-MA baselines at the same concurrency (K=1), this represents a 17.4% and 54.4% reduction in perplexity, respectively. Even at higher concurrencies (K=2), the MA system maintained a lead with 14.2% (Alpaca) and 13.7% (Dolly) lower perplexity.
Participation Fairness
Beyond accuracy, the MA system significantly improves fairness. In wireless networks, "weak" clients with poor signals are often excluded to maintain aggregation quality. By optimizing antenna positions to boost these weak links, the system increased the Jain Index (a measure of fairness) from 0.285 to 0.789 at low concurrency (K=1), meaning the system was nearly three times better at involving all clients.
Conclusion
This work demonstrates that the physical geometry of the communication interface is a powerful tool for optimizing distributed machine learning. By co-optimizing antenna positions with training hyperparameters, the proposed system overcomes traditional trade-offs between communication efficiency, model quality, and client fairness in 6G-enabled federated learning. Future research will explore the application of this technology in more complex multipath environments and investigate the long-term mechanical reliability of movable antenna systems.
Summary updated