研究方向：端侧大模型系统

Algorithm–System Co-Design

On-Device Agentic LLM System Framework

模型层
Layer I Foundation Model Pre- & Post-Training
◆ 基模架构与 MoE
紧凑基模与稀疏专家路由
◆ 能力蒸馏与对齐
指令遵循与推理链蒸馏
◆ 持续学习与 SFT
免重训持续微调 (VaLoRA)
▼ ▲
系统层
Layer II On-Device Agent Harness
▸ Runtime 调度
低延迟规划与显存协同
▸ 图 RAG & 记忆
终身时序记忆 (SmartRAG)
▸ 工具 & 扩展
云边协同工具调用底座
▼ ▲
应用层
Layer III Domain-Specific Agent Applications
◇ 情感计算 Agent
隐私共情交互 (MobileOC)
◇ 直播交互 Agent
实时流式助播与工具协同
◇ 具身智能 Agent
感知-规划-控制闭环

Layer I. 基模预训练与后训练

  • Edge Backbone & MoE
    SMoE, ISCA'26; APNet'26
  • Distillation & Alignment
  • Continual Post-training
    VaLoRA, EuroSys'25; TMC'26

Layer II. 端侧 Agent Harness

  • Agent Runtime Engine
  • Native Graph RAG
    SmartRAG, arXiv'26
  • Tool Scaling
    APNet'26

Layer III. 领域智能体应用

  • Affective Computing
    MobileOC, UbiComp'26
  • Live-Streaming Agent
  • Embodied AI Agent

For full papers, see Publications. We recruit motivated students — see Home