Redefining Information Freshness: AoGI for Generative AI in 6G Networks
Yuquan Xiao, Qinghe Du, Wenchi Cheng, George K. Karagiannidis, Arumugam Nallanathan, Mohsen Guizani
TL;DR
Real-time GenAI in 6G MEC networks requires new freshness and trust metrics beyond traditional AoI. The authors define Age of Generative Information (AoGI) and Age of Trust (AoT) to capture the effects of sampling, transmission, computation, and zero-trust verification on information timeliness and reliability. They propose optimization strategies including model partitioning, quantization/pruning, early exits, caching, and EVaR-based risk metrics (statistical AoGI) to balance freshness and security under dynamic wireless and compute resources. A demonstration shows AoGI-oriented designs can reduce average peak AoGI by about 25% compared with delay-focused approaches, highlighting practical gains for real-time GenAI services. The work also outlines open challenges in green deployment, personalization, and cross-device collaboration for MEC-enabled GenAI systems.
Abstract
Generative Artificial Intelligence (GenAI) is playing an increasingly important role in enriching and facilitating human life by generating various useful information, of which real-time GenAI is a significant part and has great potential in applications such as real-time robot control, automated driving, augmented reality, etc. There are a variety of information updating processes in real-time GenAI, and the age of information (AoI) is an effective metric for evaluating information freshness. However, due to the diversity and generativity of information in real-time GenAI, it may be incompatible to directly use existing information aging metrics to assess its timeliness. In this article, we introduce a new concept called Age of Generative Information (AoGI) to evaluate the freshness of generative information, which takes into account the information delay caused not only by sampling and transmission, but also by computation. Furthermore, since real-time GenAI services are often supported by mobile-edge-cloud (MEC) collaborative computing in 6G networks and some of the generated information is privacy sensitive, it is recommended that the identities of edge and cloud should always be verified in a zero-trust manner. We introduce the concept of Age of Trust (AoT) to characterise the decay process of their trust level. We also discuss the optimisations of these evolved information aging metrics, focusing on the impact of dynamic external conditions, including wireless environments and limited computational resources. Finally, we highlight several open challenges in providing timeliness guarantees for real-time GenAI services.