Research map / Reinforcement learning / RL from feedback
GUI-G1: Understanding R1-Zero-Like Training for Visual Grounding in GUI Agents
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Analyzes R1-Zero-like GRPO training for GUI visual grounding and proposes GUI-G1-3B with a fast thinking template, a box-size reward, and a GRPO objective without length normalization but with difficulty weighting, reaching 90.3% on ScreenSpot using 17K samples.
Problem
Understanding why R1-Zero-like RL training for GUI grounding underperforms or is hacked: long reasoning hurts grounding, hit/IoU rewards induce box-size reward hacking, and GRPO has length and difficulty biases.
Approach
Decomposes the R1-Zero-like pipeline into input template, reward function, and policy objective; replaces the thinking template with a no-thinking fast template, adds a box-size reward R_Box as a regularizer alongside hit and IoU rewards, and modifies GRPO by replacing |o_i| with a constant max token count and weighting each query by a relative-box-size difficulty coefficient.
Key results
- GUI-G1-3B reaches 90.3% average accuracy on ScreenSpot, above InfiGUI-R1-3B (87.5%).
- GUI-G1-3B reaches 37.1% on ScreenSpot-Pro, above UI-TARS-7B (35.7%) and InfiGUI-R1-3B (35.7%).
- GUI-G1-3B generates about 37-39 output tokens per example versus 107-114 for InfiGUI-R1-3B.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use GUI-G1-3B instead of InfiGUI-R1-3B | GUI grounding models · ScreenSpot | accuracy (Avg.): 87.5 -> 90.3 | Helped |
| use GUI-G1-3B instead of UI-TARS-7B | GUI grounding models · ScreenSpot-Pro | Avg.: 35.7 -> 37.1 | Helped |
| use GUI-G1-3B instead of InfiGUI-R1-3B | GUI grounding models · ScreenSpot-Pro | Avg.: 35.7 -> 37.1 | Helped |
| use longer reasoning chains | GUI grounding models · ScreenSpot | accuracy | Hurt |
| use a Fast Thinking Template without explicit deliberation | GUI grounding models · ScreenSpot | accuracy | Helped |
| optimize with the Hit-based reward | GUI grounding policy · ScreenSpot | accuracy | Helped |
| optimize with the Hit-based reward | GUI grounding policy · ScreenSpot | IoU | Hurt |
| optimize with the IoU-based reward | GUI grounding policy · ScreenSpot | IoU | Helped |
| optimize with the IoU-based reward | GUI grounding policy · ScreenSpot | accuracy | Hurt |
| add a box-size reward constraint alongside Hit and IoU rewards | GUI grounding policy · ScreenSpot | accuracy and IoU | Helped |
| use the box-size reward alone | GUI grounding policy · ScreenSpot | output format validity | Hurt |
| replace response-length normalization with Max_Tokens (ablation) | Standard GRPO · ScreenSpot | accuracy (Avg.): 82.3 -> 83.2 | Helped |
Surprising findings
- Longer chain-of-thought reasoning consistently degrades GUI grounding accuracy on ScreenSpot.
- Using the R_Box reward alone causes the model to fail to produce correctly formatted outputs.
Limitations
- Focuses only on grounding; does not cover action prediction or long-horizon planning.
- Analysis mainly covers online GRPO; other factors such as data composition and hyperparameters are not fully explored.
- Trained on a relatively small public dataset, which limits the performance ceiling.
Experimental setup
Datasets: UI-BERT OS-Atlas ScreenSpot ScreenSpot-Pro
Benchmarks: ScreenSpot ScreenSpot-Pro
Metrics: accuracy IoU average output tokens
Baselines: InfiGUI-R1-3B GUI-R1-3B GUI-R1-7B UI-R1-3B UI-TARS-7B OS-Atlas-7B UGround-7B Claude Computer Use
Builds on: GRPO DeepSeek-R1-Zero VLM-R1 InfiGUI-R1
Closest papers by meaning
- SE-GUI: Enhancing Visual Grounding for GUI Agents via Self-Evolutionary Reinforcement Learning (NeurIPS 2025 · similarity 0.93)
- UI-Ins: Enhancing GUI Grounding with Multi-Perspective Instruction as Reasoning (ICLR 2026 · similarity 0.92)
- Learn where to Click from Yourself: On-Policy Self-Distillation for GUI Grounding (NeurIPS 2026 · similarity 0.92)
- GuirlVG: Incentivize GUI Visual Grounding via Empirical Exploration on Reinforcement Learning (ICLR 2026 · similarity 0.92)
- GUI-Eyes: Tool-Augmented Perception for Visual Grounding in GUI Agents (AAAI 2026 · similarity 0.92)
- Test-Time Reinforcement Learning for GUI Grounding via Region Consistency (AAAI 2026 · similarity 0.92)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.