Yining Ye(叶奕宁)

Hi, I am a master student at THUNLP lab, Tsinghua University, advised by Zhiyuan_Liu. I’m Working on LLM/VLM Agent. I believe the future will not invent itself.

Selected Publications

2025/09: UI-TARS-2 Technical Report: Advancing GUI Agent with Multi-Turn Reinforcement Learning

We further scales ui tars to bigger model, more data, and more tools. We actually unifies MCP interface, Code interface, GUI interface, and even raw mouse/keyboard actions in one model. We run end2end Agent RL in all settings, using one value model. We scales the max iteration budgets of training with more than 1000 GPUs, and achieves SOTA performance on all agent settings. We see some possibility to use hybrid tool calls in one trajectory, and emergent ability to perform on unseen settings and tool set.

ToolLlama finally works, after 2 years, with 100x more GPU, 50x paramters, with end2end RL.

2025/01: UI-TARS: Pioneering Automated GUI Interaction with Native Agents (6k star)

We trained a native GUI agent model that solely perceives the screenshots as input and performs human-like interactions (e.g., keyboard and mouse operations). TARS achieves SOTA performance in 10+ GUI agent benchmarks evaluating perception, grounding, and GUI task execution. Notably, in the OSWorld benchmark, UI-TARS achieves scores of 24.6 with 50 steps and 22.7 with 15 steps, outperforming Claude (22.0 and 14.9 respectively)

It’s basically the inverse XAgent

2024/02: RepoAgent: An LLM-Powered Open-Source Framework for Repository-level Code Documentation Generation(EMNLP 2024)

Our goal is to create an intelligent document assistant that helps people read and understand repositories and generate documents, ultimately helping people improve efficiency and save time.

2024/01: DebugBench: Evaluating Debugging Capability of Large Language Models (ACL 2024)

We have evaluated the debugging abilities of common LLMs, and we found that open-source LLM did poor on that task

2023/11: ProAgent: From Robotic Process Automation to Agentic Process Automation

We explored using LLM to automatically generate RPA workflows, and how to use LLM as AI-data and AI-logic node in the workflow, which is called APA(Agentic Process Automation)

2023/10: XAgent (8k star)

XAgent is an open-source experimental Large Language Model (LLM) driven autonomous agent that can automatically solve various tasks. It is designed to be a general-purpose agent that can be applied to a wide range of tasks. XAgent is still in its early stages, and we are working hard to improve it.

2023/8: Large Language Model as Autonomous Decision Maker (ICLR 2025)

We Provided a novel Elo-based tree search method, connecting prior and posterior knowledge, and reaching the SOTA on the ToolBench Dataset

2023/7: Toolllm: Facilitating large language models to master 16000+ real-world apis (ICLR 2024 spotlight)

We aligned 16000+ real-world RapidAPI query, tested ChatGPT and GPT-4 to automaticaly handle real-world without human knowledge. Together, we trained Llama on the annotated data, making Llama the same function calling ability with ChatGPT

2023/4: Tool Learning with Foundation Models (CORR)

We make the first step towards general tool learning settings, testing on about 30 tasks.

Honors & Awards

Siebel-Scholar 2025
Outstanding-Graduate of Tsinghua University

Academic Background

Name	Time	Degree
Beijing National Day School	2016-2019	High school
Tsinghua University, Computer Science and Technology	2019-2023	Bachelor’s Degree
Tsinghua University, Computer Science and Technology, THUNLP Lab	2023-2026	Master’s Degree