Yuxiao Ye (叶语霄)

We proposed a multi-agent DRL framework, which consists of an intrinsic reward driven exploitation of agent’s individuality, enabling the accurate division of work, and a meta-learning based policy optimization, facilitating flexible cooperation modeling among agents.

@inproceedings{ye2023exploring,
  title={Exploring both individuality and cooperation for air-ground spatial crowdsourcing by multi-agent deep reinforcement learning},
  author={Ye, Yuxiao and Liu, Chi Harold and Dai, Zipeng and Zhao, Jianxin and Yuan, Ye and Wang, Guoren and Tang, Jian},
  booktitle={2023 IEEE 39th International Conference on Data Engineering (ICDE)},
  pages={205--217},
  year={2023},
  organization={IEEE}
}
        

We proposed a multi-agent DRL framework, with a traffic flow prediction mechanism based on spatial-temporal transformer, and a graph-based inter-agent communication method, to achieve efficient path planning for agents.

@article{ye2023qoi,
  title={QoI-Aware Mobile Crowdsensing for Metaverse by Multi-Agent Deep Reinforcement Learning},
  author={Ye, Yuxiao and Wang, Hao and Liu, Chi Harold and Dai, Zipeng and Li, Guozheng and Wang, Guoren and Tang, Jian},
  journal={IEEE Journal on Selected Areas in Communications},
  year={2023},
  publisher={IEEE}
}
        

We proposed a model-based DRL framework that primarily consists of a novel Monte Carlo Tree Search method. We further enhanced it by incorporating a spatial UAV-user correlation extraction mechanism through a relational graph convolutional network, aimed at minimizing the Age of Information (AoI) of mobile users in the mobile crowdsensing campaign.

@inproceedings{dai2022aoi,
  title={Aoi-minimal uav crowdsensing by model-based graph convolutional reinforcement learning},
  author={Dai, Zipeng and Liu, Chi Harold and Ye, Yuxiao and Han, Rui and Yuan, Ye and Wang, Guoren and Tang, Jian},
  booktitle={IEEE INFOCOM 2022-IEEE Conference on Computer Communications},
  pages={1029--1038},
  year={2022},
  organization={IEEE}
}
        

We constructed a new Text-to-SQL benchmark to mitigate overfitting in LLMs, conducted comprehensive evaluations on five Text-to-SQL sub-tasks across six LLMs, identified the distinct capabilities and limitations of LLMs, and proposed optimal in-context learning solutions tailored to each sub-task.

@article{zhang2024benchmarking,
  title={Benchmarking the Text-to-SQL Capability of Large Language Models: A Comprehensive Evaluation},
  author={Zhang, Bin and Ye, Yuxiao and Du, Guoqing and Hu, Xiaoru and Li, Zhishuai and Yang, Sun and Liu, Chi Harold and Zhao, Rui and Li, Ziyue and Mao, Hangyu},
  journal={arXiv preprint arXiv:2403.02951},
  year={2024}
}
        

We proposed an LLM-based Text-to-SQL framework, consisting of an enhancement of in-context learning and schema linking, and a cross-consistency mechanism across different models, which achieves new SOTA results on the Spider benchmark with an accuracy of 87.6%.

@article{li2024pet,
  title={PET-SQL: A Prompt-enhanced Two-stage Text-to-SQL Framework with Cross-consistency},
  author={Li, Zhishuai and Wang, Xiang and Zhao, Jingjing and Yang, Sun and Du, Guoqing and Hu, Xiaoru and Zhang, Bin and Ye, Yuxiao and Li, Ziyue and Zhao, Rui and others},
  journal={arXiv preprint arXiv:2403.09732},
  year={2024}
}