Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
GUI-G1: Understanding R1-Zero-Like Training for Visual Grounding in GUI Agents
Authors: Yuqi Zhou, Sunhao Dai, Shuai Wang, Kaiwen Zhou, Qinglin Jia, Jun Xu
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section, we introduce the experimental setup for training and evaluating our proposed GUI-G1-3B agent. We outline the implementation details, describe the training dataset and evaluation benchmarks, and provide a detailed comparison with state-of-the-art methods. |
| Researcher Affiliation | Collaboration | 1Gaoling School of Artificial Intelligence, Renmin University of China 2Huawei Noah s Ark Lab |
| Pseudocode | No | The paper describes methods and algorithms in paragraph text and through mathematical equations, but does not include any clearly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | The project repository is available at https://github.com/Yuqi-Zhou/GUI-G1. |
| Open Datasets | Yes | Our GUI-G1-3B, trained on 17K public samples with Qwen2.5-VL-3B-Instruct, achieves 90.3% accuracy on Screen Spot and 37.1% on Screen Spot-Pro. This surpasses all prior models of similar size and even outperforms the larger UI-TARS-7B, establishing a new state-of-the-art in GUI agent grounding. ... small (about 17K) set of grounding samples, showing strong performance with limited supervision from public datasets such as UI-BERT [3] and OS-Atlas [39]. |
| Dataset Splits | No | The paper describes the composition of its 17K training dataset by domain (Mobile, Web, Desktop from UI-BERT and OS-Atlas) and uses separate benchmarks (Screen Spot, Screen Spot-Pro) for evaluation. However, it does not specify explicit training/validation/test splits within the 17K samples, nor does it provide details on how the evaluation benchmarks are split for testing. |
| Hardware Specification | Yes | We conduct training on 4 NVIDIA H800 GPUs over 3 days, with a global batch size of 32 and a learning rate of 1 10 6. |
| Software Dependencies | No | Our model is built upon the Qwen2.5-VL-3B-Instruct and trained using the VLM-R1 framework [35]. The paper mentions these specific models/frameworks but does not provide version numbers for general software dependencies like Python, PyTorch, or CUDA. |
| Experiment Setup | Yes | The reward function follows the form RHit + αRIo U + βRBox, where α is set to 0.25 and β to 0.125. We conduct training on 4 NVIDIA H800 GPUs over 3 days, with a global batch size of 32 and a learning rate of 1 10 6. No KL divergence regularization is applied. Only one training epoch is required. |