[2026.5]: 🤩 Our training dataset GUI-C2-4K released on HuggingFace.
[2026.5]: 🤩 Our paper is released on arXiv.

Existing agentic reinforcement learning methods for GUI grounding have limitations at two levels. At the data level, current approaches typically treat all training samples equally, although their training value to the baseline model varies with difficulty. Overlooking this can greatly reduce training efficiency or even cause collapse. At the strategy level, existing frameworks struggle to balance the trade-off between cropping larger regions for sufficient context and smaller ones for reduced redundancy, a tension inherent to tool-augmented grounding agents. In addition, overly complex decision-making is difficult for small-parameter models and significantly increases inference time. To address these issues, at the data level, we propose GUI-D, a data mining and difficulty scoring pipeline that identifies the training-worthy samples by proper testing and assigns difficulty scores to guide subsequent training weights. At the strategy level, we propose GUI-C², which employs an area-gated coarse-to-fine refinement mechanism that progressively narrows the visual field via model-internal uncertainty signals, adaptively reserving context for large targets while amplifying precision for small ones, reinforced by improvement-aware stage rewards that ensure each refinement genuinely advances grounding. Meanwhile, we simplify the decision-making process to greatly reduce additional inference time. Finally, extensive experiments show that our method achieves state-of-the-art performance. Code will be publicly available.

GUI-D is a lightweight data curation pipeline that removes low-value GUI grounding samples and assigns each remaining sample a difficulty score for GRPO training. It consists of:

• Instruction & Length Filtering: Retain action-oriented instructions; remove short item-name prompts.
• Cross-domain Filtering: Remove web-like samples found in non-web domains via OCR detection.
• Rollout Informativeness Filtering: Remove samples where all 8 rollouts succeed or all fail; keep mixed-outcome samples.
• Difficulty Scoring: A weighted sum of 8-click success rate, prediction dispersion, localization error, target size, and parsing failures, normalized per-platform.

GUI-C² simplifies the decision-making process into predicting the size of the target bounding box, which serves as both a refinement trigger and a dense supervision signal. Key components include:

• Difficulty-Aware training weight adjustment: We incorporate sample-wise difficulty scores as interval-bounded loss weighting factors into the GRPO objective, which upweights gradients from challenging instances while limiting the maximum weight disparity to safeguard against overfitting to noisy difficulty estimates.
• Area-gated refinement: The model predicts a bounding box on the full screenshot; if the predicted area is smaller than threshold τ₁, it crops the region and predicts again (Stage 2); if still smaller than τ₂, it crops once more (Stage 3).
• No explicit thinking: Unlike other agentic frameworks, GUI-C² removes the thinking stage entirely, reducing inference time significantly.
• Improvement-aware stage rewards: Each refinement stage receives an auxiliary reward that encourages effective crops and penalizes harmful ones.

• GUI-D: A data filtering and dynamic weighting pipeline for GRPO that leverages the difficulty metric to optimize sample utilization, improving training efficiency.
• GUI-C²: An efficient agentic RL framework for coarse-to-fine GUI grounding. By formulating multi-stage cropping as bounding box prediction, it applies distinct learning behaviors to simple and hard samples, reducing decision complexity while improving inference efficiency.
• By combining dynamic difficulty weighting and coarse-to-fine cropping, our method achieves state-of-the-art performance under comparable conditions (3B) on three commonly used benchmarks, remarkably using only 4,624 training samples.