Search papers, labs, and topics across Lattice.
This research is supported by the RIE2025 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) (Award I2301E0026), administered by A*STAR, as well as supported by Alibaba Group and NTU Singapore through Alibaba-NTU Global e-Sustainability CorpLab (ANGEL). (Corresponding author: Dacheng Tao.)Shunyu Liu, Junjie Zhang, Rongcheng Tu and Dacheng Tao are with Nanyang Technological University, Singapore (e-mail: shunyu.liu@ntu.edu.sg; junjie.zhang@ntu.edu.sg; turongcheng@gmail.com; dacheng.tao@ntu.edu.sg).Wenkai Fang, Yang Zhou, Kongcheng Zhang, and Mingli Song are with the College of Computer Science and Technology, Zhejiang University, China (e-mail: wenkfang@zju.edu.cn; imzhouyang@zju.edu.cn; zhangkc@zju.edu.cn; brooksong@zju.edu.cn).Zetian Hu is with the School of Aerospace Engineering, Tsinghua University, China (e-mail: huzt22@mails.tsinghua.edu.cn).Ting-En Lin, Fei Huang, and Yongbin Li are with the Tongyi Lab, Alibaba Group, China (e-mail: ting-en.lte@alibaba-inc.com; f.huang@alibaba-inc.com; shuide.lyb@alibaba-inc.com)
Tsinghua AI3
0
6
0
AI agents can now learn durable skills instead of constantly "reinventing the wheel," thanks to SkillNet's infrastructure for creating, evaluating, and connecting AI skills at scale.
Training web agents in a simulator can now match real-world performance: Qwen3-14B, fine-tuned with WebWorld-synthesized trajectories, rivals GPT-4o on WebArena.
Function calling gets a serious upgrade: a new reward model and inference scaling technique boosts performance by focusing on the *process* of tool use, not just the outcome.
