Movable Antenna Enhanced Federated Fine-Tuning of Large Language Models via Hybrid Client Selection Optimization

Large language model (LLM) training in 6G networks faces stringent latency and energy constraints while operating over bandwidth-limited wireless links. A commonly adopted workflow separates training into a centralized pre-training phase and a federated fine-tuning phase on domain-specific data; however, over-the-air (OTA) gradient aggregation during fine-tuning remains vulnerable to fading and interference. This study explores the integration of movable antennas (MAs), whose element positions can be reconfigured in real time, to mitigate such channel impairments. An auxiliary channel representation embeds transmit power terms in the effective gain, thereby removing explicit power-control variables. We derive the convergence bound that determines the relationship between the final fine-tuning loss to OTA noise and the distribution shift between the two data stages, measured via the Wasserstein distance. These findings lead to a mixed integer, nonconvex resource allocation problem that jointly determines the numbers of global rounds, CPU frequencies, mini-batch sizes, positions of MAs, and beamformers under latency-energy constraints. We propose a hybrid successive convex approximation (SCA) and penalty dual decomposition (PDD) algorithm to solve the problem. Experiments with the OpenLLaMA v2 model with 3 billion parameters demonstrate up to $95%$ faster convergence and over $90%$ lower total energy consumption relative to the leading wireless federated learning baselines, underscoring the promise of MAs-assisted federated LLM fine-tuning for 6G edge intelligence.