ByteDance releases UI-TARS-2

Background

UI-TARS-2 is the next model in ByteDance Seed's GUI-agent line, following UI-TARS-1.5. A GUI agent operates software through its screen, so each step means reading what is on the screen and choosing a click, a keystroke or a scroll. Training such an agent is a multi-step problem: the result of an early action affects every later one.

What happened

ByteDance Seed posted the UI-TARS-2 technical report to arXiv on 2 September 2025 and announced the release on GitHub on 4 September 2025.^[1][2]

The model was trained with multi-turn reinforcement learning, a method in which the model is rewarded or penalised for the outcome of a whole sequence of actions rather than for each step alone. The authors described UI-TARS-2 as an all-in-one upgrade over UI-TARS-1.5.^[1][2]

Key results

The report gives these scores: 88.2 on Online-Mind2Web, 47.5 on OSWorld, 50.6 on WindowsAgentArena and 73.3 on AndroidWorld.^[1]

Across 15 games, the model reached about 60% of human-level performance, with a mean normalised score of 59.8.^[1]

The authors said UI-TARS-2 outperformed Claude and OpenAI agents on GUI benchmarks. On LMGame-Bench, a benchmark built from games, it was competitive with OpenAI o3.^[1]

Reinforcement learning is a training method in which a model learns from the reward its actions earn, rather than only from examples of correct answers. For a multi-step agent the reward often arrives only at the end of a task, so the model has to work out which early actions lead to success. This is harder than training a model to answer a single question, because a mistake at one step changes everything that follows.

Multi-turn training means the agent goes through many rounds of observing and acting before a task ends, so errors can build up if the training does not account for the whole sequence.

Significance

The game results test a different skill from the desktop and Android benchmarks: they require the agent to keep track of changing scenes over many steps.