Technical Report 2025 Technical report

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

A native GUI agent trained for multi-turn interaction across desktop, mobile, and game environments.

ByteDance Seed

The report is presented by ByteDance Seed. Its full individual author list includes Qinyu Luo; the team credit here follows the report cover.

UI-TARS-2 data flywheel connecting model training, annotation, and data synthesis. — Fig. 3 — Original UI-TARS-2 data flywheel.Original source ↗

01 / The question

How can GUI agents learn reliably from long interactions in realistic software environments?

02 / The approach

UI-TARS-2

UI-TARS-2 combines a data flywheel, stabilized multi-turn reinforcement learning, and a unified sandbox for large-scale rollouts. The technical report studies GUI operation, game interaction, and transfer to broader agent tasks.

Read the full paper

03 / Citation

Cite this work

BibTeX

@misc{wang2025uitars2,
  title = {UI-TARS-2 Technical Report: Advancing GUI Agent with Multi-Turn Reinforcement Learning},
  author = {Wang, Haoming and Zou, Haoyang and Song, Huatong and Feng, Jiazhan and Fang, Junjie and Lu, Junting and Liu, Longxiang and Luo, Qinyu and others},
  year = {2025},
  eprint = {2509.02544},
  archivePrefix = {arXiv},
  primaryClass = {cs.AI},
  url = {https://arxiv.org/abs/2509.02544}
}

Select the citation to copy it manually.

04 / Versions & records

Publication record

3 Google Scholar records

READ HERE · 30 PAGES

Full paper

Document

Loading document…

Scroll through every page. Select text to copy, or use search to find a passage.

More research

View all

How can GUI agents learn reliably from long interactions in realistic software environments?

UI-TARS-2

Cite this work

Publication record

Full paper

More research

TTExplore↗

Agent-World↗

TTExplore

Agent-World