Intelligent Music Streaming Scheduling and QoE Optimization in 6G Wireless Networks Using Large-Scale Models

Abstract

In this paper, we propose LM-QoEStream, a novel framework that integrates large-scale language models (LLMs) with reinforcement learning-based streaming scheduling to optimize music delivery under dynamic wireless conditions. Specifically, we design a prompt-driven Quality of Experience (QoE) prediction module that transforms heterogeneous user, content, and network features into structured natural language prompts, enabling the LLM to infer fine-grained user satisfaction scores. These scores are then used as rewards in a Soft Actor-Critic (SAC) reinforcement learning (RL) controller that dynamically adjusts streaming parameters such as bitrate and buffer strategies. Extensive experiments conducted on simulated 5G/6G networks with real-world content and user interaction traces demonstrate that LM-QoEStream significantly outperforms baseline methods in terms of average QoE, stall ratio, bitrate adaptation accuracy, and fairness. Ablation studies further confirm the complementary strengths of both the LLM-based perception model and the learning-based decision module. The proposed approach offers a scalable, generalizable, and user-centric solution for next-generation music streaming systems.