Selected work

Representative Papers

What Visual Generators Need from Teachers: Rethinking Representation Alignment

Sep 28, 2026

This study addresses the reliance on empirical trial-and-error for teacher layer selection in knowledge distillation for diffusion models by proposing the RARE method. By revealing that student networks populate features in a bottom-up, layer-by-layer manner, this work defines a "recoverability gap" to replace conventional similarity metrics. This formulation enables adaptive alignment target selection and dynamic training loss weighting without trial-and-error. Evaluated on ImageNet, RARE reduces the FID to 4.46 with guidance, outperforming baselines such as REPA while decreasing GPU training time by 14%. Furthermore, the proposed approach demonstrates robust generalization across datasets of varying scales.

0 citationsRead paper

AISC deployment in dynamic UAV-assisted MEC network: a reinforcement learning method based on heterogeneous graph attention neural network

Jun 04, 2026

This study addresses the challenge of deploying AI service chains (AISCs) in dynamic unmanned aerial vehicle–assisted mobile edge computing (UMEC) networks, where high topological dynamics, complex interdependencies among virtual network functions (VNFs), and trade-offs between energy consumption and load balancing hinder minimization of service completion time. To tackle this, the work proposes a dual deep attention Q-network method that integrates heterogeneous graph attention into a deep reinforcement learning framework. By modeling heterogeneous nodes and links in the drone network and leveraging attention mechanisms to adaptively focus on critical resources, the approach enables end-to-end optimized AISC deployment. Experimental results demonstrate that the proposed method significantly outperforms existing baselines in terms of service completion time, success rate, load balancing, and energy efficiency.

0 citationsRead paper

Availability-Aware and Efficiency-Driven AI Service Chain Provisioning in Multi-Domain Edge Intelligence Cloud

Jun 03, 2026

This study addresses the challenge of efficiently and reliably deploying AI service chains in multi-domain edge intelligent clouds, where heterogeneous environments, resource constraints, and partial observability of system states hinder performance. To tackle this problem, the authors formulate it as a partially observable stochastic game and propose a graph-temporal dual network (GTDN)-based multi-agent collaborative optimization method. This approach uniquely integrates network topology and temporal dependencies among services into a unified model, enabling joint optimization of cost, latency, and availability. Experimental results demonstrate that the proposed method significantly outperforms existing baselines across diverse network topologies and edge node scales, achieving substantial improvements in overall performance.

0 citationsRead paper

Recent publications

Latest Papers

What Visual Generators Need from Teachers: Rethinking Representation Alignment

Sep 28, 2026

This study addresses the reliance on empirical trial-and-error for teacher layer selection in knowledge distillation for diffusion models by proposing the RARE method. By revealing that student networks populate features in a bottom-up, layer-by-layer manner, this work defines a "recoverability gap" to replace conventional similarity metrics. This formulation enables adaptive alignment target selection and dynamic training loss weighting without trial-and-error. Evaluated on ImageNet, RARE reduces the FID to 4.46 with guidance, outperforming baselines such as REPA while decreasing GPU training time by 14%. Furthermore, the proposed approach demonstrates robust generalization across datasets of varying scales.

0 citationsRead paper

AISC deployment in dynamic UAV-assisted MEC network: a reinforcement learning method based on heterogeneous graph attention neural network

Jun 04, 2026

This study addresses the challenge of deploying AI service chains (AISCs) in dynamic unmanned aerial vehicle–assisted mobile edge computing (UMEC) networks, where high topological dynamics, complex interdependencies among virtual network functions (VNFs), and trade-offs between energy consumption and load balancing hinder minimization of service completion time. To tackle this, the work proposes a dual deep attention Q-network method that integrates heterogeneous graph attention into a deep reinforcement learning framework. By modeling heterogeneous nodes and links in the drone network and leveraging attention mechanisms to adaptively focus on critical resources, the approach enables end-to-end optimized AISC deployment. Experimental results demonstrate that the proposed method significantly outperforms existing baselines in terms of service completion time, success rate, load balancing, and energy efficiency.

0 citationsRead paper

Availability-Aware and Efficiency-Driven AI Service Chain Provisioning in Multi-Domain Edge Intelligence Cloud

Jun 03, 2026

This study addresses the challenge of efficiently and reliably deploying AI service chains in multi-domain edge intelligent clouds, where heterogeneous environments, resource constraints, and partial observability of system states hinder performance. To tackle this problem, the authors formulate it as a partially observable stochastic game and propose a graph-temporal dual network (GTDN)-based multi-agent collaborative optimization method. This approach uniquely integrates network topology and temporal dependencies among services into a unified model, enabling joint optimization of cost, latency, and availability. Experimental results demonstrate that the proposed method significantly outperforms existing baselines across diverse network topologies and edge node scales, achieving substantial improvements in overall performance.

0 citationsRead paper