![[2602.22190] GUI-Libra: Training Native GUI Agents to Reason and Act with Action-aware Supervision and Partially Verifiable RL](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.22190] GUI-Libra: Training Native GUI Agents to Reason and Act with Action-aware Supervision and Partially Verifiable RL
Summary
The paper presents GUI-Libra, a novel training approach for native GUI agents that enhances reasoning and action capabilities through action-aware supervision and partially verifiable reinforcement learning.
Why It Matters
This research addresses significant gaps in the development of open-source GUI agents, particularly in their ability to perform complex navigation tasks. By introducing a curated dataset and innovative training techniques, the study aims to improve the effectiveness of GUI agents, which is crucial for advancing AI applications in user interface automation and interaction.
Key Takeaways
- GUI-Libra improves the performance of native GUI agents in long-horizon navigation tasks.
- The study introduces a curated dataset of 81K GUI reasoning instances to enhance training.
- Action-aware supervised fine-tuning (SFT) is proposed to better align reasoning with actions.
- KL regularization is identified as critical for stabilizing reinforcement learning under partial verifiability.
- The results indicate that effective data curation can enhance task-solving capabilities without extensive online data collection.
Computer Science > Machine Learning arXiv:2602.22190 (cs) [Submitted on 25 Feb 2026] Title:GUI-Libra: Training Native GUI Agents to Reason and Act with Action-aware Supervision and Partially Verifiable RL Authors:Rui Yang, Qianhui Wu, Zhaoyang Wang, Hanyang Chen, Ke Yang, Hao Cheng, Huaxiu Yao, Baoling Peng, Huan Zhang, Jianfeng Gao, Tong Zhang View a PDF of the paper titled GUI-Libra: Training Native GUI Agents to Reason and Act with Action-aware Supervision and Partially Verifiable RL, by Rui Yang and 10 other authors View PDF HTML (experimental) Abstract:Open-source native GUI agents still lag behind closed-source systems on long-horizon navigation tasks. This gap stems from two limitations: a shortage of high-quality, action-aligned reasoning data, and the direct adoption of generic post-training pipelines that overlook the unique challenges of GUI agents. We identify two fundamental issues in these pipelines: (i) standard SFT with CoT reasoning often hurts grounding, and (ii) step-wise RLVR-tyle training faces partial verifiability, where multiple actions can be correct but only a single demonstrated action is used for verification. This makes offline step-wise metrics weak predictors of online task success. In this work, we present GUI-Libra, a tailored training recipe that addresses these challenges. First, to mitigate the scarcity of action-aligned reasoning data, we introduce a data construction and filtering pipeline and release a curated 81K GUI reasoning datase...