Joint Inference-Aware and Resource Allocation Optimization for Drift Adaptation in Edge Intelligence

The deployment of large language models (LLMs) on mobile devices holds great potential for enabling intelligent applications, but it also poses significant challenges in inference accuracy, latency, and model robustness. Pre-trained LLMs suffer inference accuracy degradation due to mobile data drift in dynamically changing mobile scenarios, and frequent data uploads for retraining bring substantial communication overhead. Further, resource-constrained mobile edge networks complicate effective system optimization. To address these challenges, we propose AdaInfer, an edge-end collaborative system, to optimize inference accuracy and resource allocation. Specifically, AdaInfer relies on edge servers to continuously retrain and re-compress LLMs with freshly collected data from mobile devices. To overcome resource constraints, we formulate an optimization problem that balances inference accuracy and system efficiency, which is efficiently solved by a self-adaptive global best harmony search (SGHS) algorithm. Experimental results show that AdaInfer achieves high quality inference and efficient resource allocation in dynamic mobile environments.