XXXinyi Xu

Xinyi Xu

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Xinyi Xu has a total of 19 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Joint Inference Offloading and Model Caching for Small and Large Language Model Collaboration

    by Xinyi Xu, Gang Feng, Yijing Liu, Shuang Qin, Jian Wang, Yunxiang Wang

    Large Language Models (LLMs), with advanced content creation and inference capabilities, can provide immersive intelligent services to users in mobile edge networks. However, the increasing demand for real-time artificial intelligence (AI) applications aggravates the limitations of cloud-based LLMs due to the long response time. Meanwhile, Small Language Models (SLMs), which are cost-effective and locally deployable for terminal devices, can serve as an efficient supplement to LLMs for performing latency-sensitive tasks with lower generalization capability. Due to the resource constraints of edge networks and the diverse requirements of user tasks, it is critical to design an inference framework that effectively coordinates the deployment and collaboration of LLMs and SLMs. In this paper, we propose an LLM-SLM collaborative inference (LSCI) scheme under a mobile edge computing (MEC) architecture, which jointly decides where to cache models and how to offload inference tasks to balance latency, accuracy, and resource costs. To optimize inference performance subject to resource constraints, we jointly solve the inference task offloading and model caching problem in LSCI scheme. Specifically, we employ deep reinforcement learning (DRL) to select highly popular SLMs to be cached on the edge server, and distributed belief propagation technique to solve the associated inference task offloading issue. Numerical results show that the proposed LSCI scheme can achieve significant performance gain in terms of inference performance when compared with a number of baseline solutions.


  • Split Learning Based Cloud-Edge-End Collaborative Model Training in Heterogeneous Networks

    by Jian Wang, Gang Feng, Yi-Jing Liu, Xinyi Xu, Lei Cheng, Wei Jiang, Li Ping Qian

    Large language models (LLMs) demonstrate significant potential for enabling intelligent endogenous networks owing to their amazing intelligence level. LLMs are currently deployed on cloud servers as their vast parameter scales introduce a substantial computational burden to pre-training and inference processes. This deployment paradigm faces severe challenges, including insufficient personalization, high inference latency, and privacy concerns, as an increasing number of mobile users enjoy the LLM services. To enhance the personalization of pre-trained LLMs, exploiting massive local datasets distributed across edge devices to fine-tune LLMs is essential. In this paper, we propose a split learning-based cloud-edge-end collaborative training framework (SCCT) to harness the abundant computational resources of cloud and edge servers while exploiting massive distributed datasets on edge devices for LLM fine-tuning. In SCCT, an LLM with a parameter-efficient fine-tuning module is deployed on the cloud server, while a small-scale language model (SLM) is deployed across each edge device participating in SCCT and its associated edge server in the manner of split learning. SLM and LLM collaborate in a serial manner for training and inference. To optimize the collaboration efficiency, we formulate a mixed integer nonlinear programming problem to minimize the training latency of SCCT, considering the resource heterogeneity of edge devices and network dynamics. To solve this problem, a two-timescale optimization algorithm is proposed to determine the optimal split points on the large timescale and the allocation of communication and computational resources of the edge server on the small timescale. Numerical results demonstrate that the proposed two-timescale optimization algorithm assisted SCCT outperforms the state-of-the-art LLMs deployment and training paradigm in terms of training efficiency and inference performance.


  • Knowledge Sharing Based Fine-Tuning for Large Pre-Trained Model in Wireless Networks

    by Yunxiang Wang, Gang Feng, Yijing Liu, Shuang Qin, Jianhong Zhou, Xinyi Xu

    Large pre-trained models (LPMs) provide essential technical support for downstream Artificial Intelligence (AI) tasks spawning under the intelligent evolution of wireless networks. Using parameter-efficient fine-tuning (PEFT) technology, LPM can effectively capture task-specific PEFT knowledge and generalize to new tasks. However, the inherent system heterogeneity of wireless networks and data across user equipments (UEs) cause the heterogeneous PEFT models. This in turn brings difficulties in combining the PEFT knowledge of users, and thus degrades the global generalization capability of LPM. Meanwhile, additional computational and communication costs will be incurred if conventional heterogeneous model training methods are used for LPM to capture the global PEFT knowledge. Therefore, effective PEFT knowledge sharing becomes crucial for exploiting the global generalization capability of LPM. In this paper, we propose a PEFT knowledge sharing based distributed fine-tuning (DFT) scheme, by aggregating heterogeneous PEFT models. In DFT, the PEFT knowledge of users is captured with low-rank adaptive (LoRA) model, where the LoRA models can be heterogeneous for different users. On this basis, we further design an aggregation and iterative learning strategy for the heterogeneous LoRA models to capture the global PEFT knowledge from downstream users. Meanwhile, users perform iterative learning based on individual reinitialized local LoRA models, so as to avoid additional communication cost in using a global LoRA model. We theoretically analyze the convergence bound of our proposed DFT scheme. The numerical results verify the effectiveness of our proposed DFT scheme in enhancing the global generalization capability of LPM while significantly reducing communication cost.


  • A Survey of Integrating Generative Artificial Intelligence and 6G Mobile Services: Architectures, Solutions, Technologies and Outlooks

    by Yi-Jing Liu, Hongyang Du, Xinyi Xu, Ruichen Zhang, Gang Feng, Bin Cao, Dusit Niyato, Dong In Kim, Abbas Jamalipour, Khaled Ben Letaief, et al.

    Generative artificial intelligence (GenAI) is rapidly driving a new phase of artificial intelligence revolution, marked by various applications such as ChatGPT, Sora and DeepSeek. With powerful capabilities in content generation, personalization and user interaction, GenAI can drive significant enhancements in various mobile services to satisfy individual user preferences. Meanwhile, mobile services offer substantial support for accessibility, functionality and intelligence of GenAI applications. However, current research on integrating GenAI and mobile services remains nascent, lacking necessary theoretical and technical support. In this survey, we break down a new vision to guide research on integrating GenAI and mobile services, aiming to improve 6G intelligence levels and enable 6G mobile networks to support the integration of GenAI and mobile services effectively. We begin by examining the necessity of integrating GenAI with 6G mobile services and exploring potential learning architectures of GenAI models within mobile networks. Next, we analyze potential solutions from two perspectives: how 6G mobile networks empower GenAI and how GenAI enables mobile services. We subsequently explore a promising use case, along with essential features and techniques, to facilitate the integration of GenAI and mobile services. Finally, we discuss future research directions, aiming at realizing intelligent 6G networks.


  • Joint UAV Deployment and Resource Allocation: A Personalized Federated Deep Reinforcement Learning Approach

    by Xinyi Xu, Gang Feng, Shuang Qin, Yijing Liu, Yao Sun

    Unmanned aerial vehicles (UAVs) are capable of serving as aerial base stations (BSs) for providing dynamic coverage and connectivity extension for the sixth-generation (6G) wireless networks. While flexibility is provided, the deployment of the UAV swarms and the associated resource allocation become rather challenging due to the dynamic nature of UAVs and difficulty in obtaining global user information. In this paper, we propose an adaptive and flexible joint UAV deployment and resource allocation (JUDRA) scheme by exploiting personalized federated deep reinforcement learning, called PFRL, with aim to maximize the long-term network throughput while enforcing user privacy and adapting to time-varying network states. To allow UAVs to make real-time decisions on resource allocation and position adjustment based on local observations while achieving a global optimal solution, a deep reinforcement learning (DRL) algorithm is adopted in the federated learning framework in PFRL. Specifically, we use DRL to train a local model and a personalized model on UAVs, and employ a two-level parameter aggregation scheme on a leading UAV to form a global model. The personalized model can adapt to changing environments, while exploiting the generalization of global model to accelerate the learning convergence. Numerical results show that the proposed PFRL scheme can achieve significant performance gain in terms of network throughput and convergence in comparison with some state-of-art solutions.