Network-aided Efficient Large Language Model Services With Denoising-inspired Prompt Compression

Large Language Models (LLMs) have demonstrated remarkable capabilities in various tasks, leading to their increasing adoption in diverse services delivered through wireless networks. There is a growing trend toward longer prompts to better leverage LLMs' capabilities and address difficult tasks. However, longer prompts not only increase data transmission costs across wireless transmission but also require more computing resources and processing time, impacting the overall system efficiency and user experience. To address this challenge, we propose Joint Power and Prompt Optimization (JPPO), a framework that combines Small Language Model (SLM)-based prompt compression with wireless power allocation optimization. By deploying SLM at edge devices for prompt compression and employing Deep Reinforcement Learning (DRL) for joint optimization of compression ratio and transmission power, JPPO effectively balances service quality with resource efficiency. Furthermore, inspired by denoising diffusion models, we design a denoising-inspired prompt compression approach that iteratively compresses prompts by gradually removing non-critical information. Experimental results demonstrate that our framework achieves high service fidelity while optimizing power usage in wireless LLM services, reducing the total service response time. With our DRL-based JPPO, the framework maintains fidelity comparable to the no-compression baseline while still achieving a 17% service time reduction through adaptive compression. When prioritizing compression, our framework achieves up to 16x compression ratio while maintaining acceptable fidelity (within 30% reduction). Compared to no compression, baseline single-round compression with a 16x compression ratio reduces the system total response time by approximately 42.3%, while the denoising-inspired method achieves a 46.5% service time-saving.