STEVE-R1: Towards Long Reasoning Computer-use Agents

We evaluate the performance of the STEVE-R1 agent on both in-domain WindowsAgentArena (Windows 11 OS) and out-of-domain OSWorld (Ubuntu OS) benchmarks. The evaluation involves 16 attempts per task, with task completion rates recorded as the primary metric. In the in-domain Windows 11 setting, the STEVE-R1 agent demonstrated a 14% higher task completion rate compared to the previous open-source state-of-the-art model, UI-TARS-7B-DPO. Furthermore, in the out-of-domain Ubuntu OS environment, where STEVE-R1 was not explicitly trained, it still achieved a 7% higher task completion rate than UI-TARS-7B-DPO.

Release

Currently only the SFT STEVE-R1 model with step-verified training data is released. RL tunning is in progress.
🔥 An improved version STEVE-R1 is released with long reasoning ability and long image context. We extend the model context length to 128K with at most 32 screenshot inputs for a single task. The model response length is greatly improved with deepseek-R1 distillation, see the examples. We release the training data, models, and evaluation trajectories.
We release the paper of STEVE: Step Verification Pipeline for Computer-use Agent Training. We propose a single-frame computer-use 7B agent trained with SFT & step-verified KTO.

Performance

Method	WinAgentArena	OSWorld
UI-TARS-7B-DPO (20 steps)	15.4 ± 1.6	14.6 ± 1.0
UI-TARS-7B-DPO (40 steps)	17.8 ± 1.3	15.7 ± 0.8
UI-TARS-7B-DPO (60 steps)	19.3 ± 1.6	16.2 ± 1.0
Our Model
STEVE-R1-SFT (20 steps)	17.5 ± 2.0	9.6 ± 1.1（OOD）
STEVE-R1-SFT (40 steps)	20.1 ± 2.2	11.5 ± 1.2 (OOD)
STEVE-R1-SFT (60 steps)	22.3 ± 2.1	12.8 ± 1.2 (OOD)
Multiple Trials
UI-TARS-7B-DPO (20 steps, pass@16)	33.1	25.2
STEVE-R1-SFT (20 steps, pass@16)	46.8	31.4 (OOD)

Trajectory Data

Evaluation trajectories on WindowsAgentArena and OSWorld: https://huggingface.co/datasets/Fanbin/waa_steve_trajectories

Project page: https://github.com/FanbinLu/STEVE-R1

Paper: https://arxiv.org/abs/2503.12532