Optimizing System Performance for Autonomous Recovery in AI-Native Networks Using Large Language Models

Future networks are expected to be AI-native, where AI is no longer an external tool but an integral element of the network infrastructure, with a complete AI lifecycle management. This lifecycle spans data collection, model training, inference, and evaluation, and must be coordinated alongside network infrastructure and application performance to maintain reliable end-to-end system behavior. However, such deep integration introduces new challenges, particularly in managing performance degradation caused by complex dependencies across lifecycle components in the system level. Traditional optimization methods, which isolate AI components as well as their interactions with networks and applications, fail to provide system-level assurance in AI-native environments. To address this, we propose a system-level performance optimization framework based on Large Language Models (LLMs) as autonomous reasoning agents within an Agentic AI paradigm. Our approach introduces a two-phase learning strategy: (1) fine-tuning the LLM with knowledge from AI solutions, networks, and applications to enable autonomous identification of performance-affecting factors, especially performance bottlenecks; and (2) real-world refinement using interaction-based learning and a reward-driven alignment mechanism that ties LLM decisions to application performance improvements. By embedding embodied intelligence into the network operation, the framework enables adaptive and lifecycle-aware performance recovery. We validate our method on a real-time 3D reconstruction task within an AI-native network, achieving a 94.5% performance recovery rate with minimal correction latency. The results demonstrate the effectiveness of LLM-powered agentic reasoning and adaptation in maintaining system performance and mitigating performance degradation in complex AI-native networking scenarios.