Abstract
Over-the-air federated learning (OTA FL) provides superior spectral efficiency and reduced communication overhead, establishing it as an effective approach for pushing intelligence to the network edge. Yet, in the era of generative AI, training distributed large models (e.g., large language models) with typically billions of parameters across edge networks is severely limited by communication bottlenecks. In this letter, we introduce ZOTA, a full zeroth-order optimization-based federated large model tuning framework with OTA computations, which bypasses the first-order optimization and addresses the dilemma between communication bottlenecks and the huge amount of large model parameters. We demonstrate that the proposed framework has extremely high communication and memory efficiency with guaranteed theoretical convergence under noisy OTA aggregation.