<aside> 💡 近年来，人工智能已逐渐在日常生活、工业生产中发挥着至关重要的作用。然而，这些人工智能应用背后的深度学习模型真实世界中部署和推理的时候，会面临推理环境与训练环境之间环境异构的挑战，主要包括，计算资源的差异、数据分布的差异等，这极大的限制了智能应用的低延迟、高性能的需求。

In recent years, artificial intelligence has been playing an increasingly crucial role in daily life and industrial production. However, when deploying and running deep learning models behind these AI applications in the real world, they face challenges of environmental heterogeneity between inference and training environments, mainly including differences in computational resources and data distribution, which greatly limits the low-latency and high-performance requirements of intelligent applications.

这个页面总结了清华大学深圳国际研究生院王智教授课题组在模型部署与微调小组方向已完成的课题，并由模型部署与微调小组小组维护。若对我们的工作感兴趣，欢迎通过以下方式联系我们：

This page summarizes the completed projects of Professor Zhi Wang's research group at Tsinghua University Shenzhen International Graduate School in the direction of model deployment and fine-tuning, and is maintained by the Model Deployment and Fine-tuning Group. If you are interested in our work, please feel free to contact us through the following methods:

</aside>

✉️ 联系方式：

• 王智 副教授（清华大学深圳国际生院）：

  • 个人主页：https://www.mmlab.top/
  • 邮箱：[email protected]

• 姜婧妍 ：

  • 邮箱： [email protected]

  Contact Information:

  • Zhi Wang, Associate Professor (Tsinghua Shenzhen International Graduate School):
    • Personal Website:https://www.mmlab.top/
    • Email: [email protected]
  • Jingyan Jiang:
    • Email: [email protected]

研究方向概览

本小组的具体研究方向分为两个部分：

• 模型泛化与持续学习

  • 测试时模型自适应推理（Test-time Adaptation）： 测试时间使用无监督的方式对模型进行在线更新，提升模型泛化性能，涉及小模型、大模型、多模态、具身智能场景…
  • LLM泛化与微调：解决 LLM 中的数据漂移和幻觉问题

• 模型高效部署与推理优化

  • 模型部署与资源调度：根据模型性能和资源消耗设计模型部署策略，在资源消耗和性能中实现平衡。
  • 大小模型联合推理：联合大模型与小模型共同实现推理任务，在资源消耗和性能中实现平衡。

  Research Direction Overview

  Our group's specific research directions are divided into two parts:

  • Model Generalization and Continuous Learning
    • Test-time Adaptation: Using unsupervised methods to update models online during testing to improve model generalization performance, involving small models, large models, multimodal, and embodied AI scenarios...
    • Federated Fine-tuning (Fed-TTA): Leveraging knowledge from multiple models to enhance large model generalization performance during testing.
  • Efficient Model Deployment and Inference Optimization
    • Model Deployment and Resource Scheduling: Designing model deployment strategies based on model performance and resource consumption to achieve balance between resource usage and performance.
    • Joint Large and Small Model Inference: Combining large and small models to accomplish inference tasks while balancing resource consumption and performance.

近期新闻（Recent News）

News！ 2025.02.28，恭喜我们的论文 COSMIC 被 CCF-A CVPR 2025 接收！

News！ 2025.03.03， 恭喜我们的论文LLM4Band 被 CCF-B NOSSDAV 2025 接收！

News！ 2025.03.21， 恭喜我们的论文DATTA被 CCF-B ICME 2025 接收！

近期工作（Recent Work）

模型泛化与持续学习（Model Generalization and Continuous Learning:）：

New！ 面向具身智能场景的持续学习（Continuous Learning for Visual Tasks）

New！面向大模型场景的持续学习（Continuous Learning for LLMs）