DRL-Based Energy-Aware Split Inference with Adaptive Compression for AI-based UE Apps in 6G Networks

While 6G moves towards native AI, the end device or User Equipment (UE) remains constrained by limited battery and compute resources, hindering the local execution of intensive AI inference pipelines. Split AI partitions inference workloads between the device and the network, reducing on-device computation and energy consumption. For mobile users, transmitting intermediate data can incur significant communication cost; while compression alleviates this overhead, it may degrade model performance. Existing solutions largely overlook the interplay among device energy constraints, user mobility, and model performance under compression. To address these challenges, we propose ENSPLIT+,a dynamic split inference framework with device-side compression for AI-based UE applications in 6G networks. We envision an AI agent at the application layer that employs ENSPLIT+ to jointly optimize inference latency, device energy consumption, and model performance under dynamic mobility, resource, energy, and application constraints. ENSPLIT+ leverages Deep Reinforcement Learning (DRL) decision making, which adaptively selects split points and compression levels. Extensive evaluations in a simulated environment with mobile users demonstrate that ENSPLIT+ reduces inference latency by up to 78% and device energy consumption by up to 60% compared to baseline split inference policies, while maintaining inference quality. These results show that ENSPLIT+ effectively balances model performance and communication efficiency, demonstrating the efficacy of DRL for AI agent planning in 6G applications.