UI-Zoomer: Uncertainty-Driven Adaptive Zoom-In for GUI Grounding Fei Tang1,∗ , Bofan Chen1,∗ , Zhengxi Lu1 , Tongbo Chen1 , Songqin Nong2 , Tao Jiang2 , Wenhao Xu2 , Weiming Lu1 , Jun Xiao1 , Yueting Zhuang1 , Yongliang Shen1† 1
arXiv:2604.14113v1 [cs.CV] 15 Apr 2026
∗
Zhejiang University, 2 Ant Group Equal Contribution, † Corresponding authors
GUI grounding, which localizes interface elements from screenshots given natural language queries, remains challenging for small icons and dense layouts. Test-time zoom-in methods improve localization by cropping and re-running inference at higher resolution, but apply cropping uniformly across all instances with fixed crop sizes, ignoring whether the model is actually uncertain on each case. We propose UI-Zoomer, a training-free adaptive zoom-in framework that treats both the trigger and scale of zoom-in as a prediction uncertainty quantification problem. A confidence-aware gate fuses spatial consensus among stochastic candidates with token-level generation confidence to selectively trigger zoom-in only when localization is uncertain. When triggered, an uncertainty-driven crop sizing module decomposes prediction variance into inter-sample positional spread and intra-sample box extent, deriving a per-instance crop radius via the law of total variance. Extensive experiments on ScreenSpot-Pro, UI-Vision, and ScreenSpot-v2 demonstrate consistent improvements over strong baselines across multiple model architectures, achieving gains of up to +13.4%, +10.3%, and +4.2% respectively, with no additional training required. Date: April 16, 2026 Project Page: https://zju-real.github.io/UI-Zoomer Code: https://github.com/ZJU-REAL/UI-Zoomer Correspondence: [email protected]
1
Introduction
Grounding natural language instructions to interface elements is a fundamental capability for autonomous GUI agents Gou et al. (2024); Tang et al. (2025a,c); Xu et al. (2025); Hong et al. (2024); Lin et al. (2024); Jiang et al. (2025); Yang et al. (2023). Despite significant progress through supervised fine-tuning and reinforcement learning Qin et al. (2025); Xu et al. (2024); Xie et al. (2025); Yuan et al. (2025); Gu et al. (2025), models still fail systematically on small icons and dense layouts in complex interfaces Li et al. (2025). A natural remedy is test-time zoom-in scaling: crop a region of the screenshot and re-run the model at higher effective resolution Wu et al. (2025a); Luo et al. (2025); Nguyen (2024); Lee et al. (2025).While this paradigm has shown clear promise for fine-grained GUI localization Wu et al. (2025a); Luo et al. (2025), a more fundamental question remains unaddressed: which instances actually need zoom-in, and how much should we zoom? Existing zoom-in methods share two fundamental limitations. First, they apply cropping indiscriminately: Wu et al. (2025a) zooms in unconditionally on every sample with a fixed scaling factor, while Luo et al. (2025) triggers zoom-in only upon execution errors, with no regard to whether the model is actually uncertain on the instance at hand. We show empirically that unconditional zoom-in on ScreenSpot-v2 degrades accuracy below the direct prediction baseline while significantly increasing latency (Table 1), as easy cases lose the global context the model was already exploiting. Second, all existing methods fix the crop window to a predetermined ratio Wu et al. (2025a); Luo et al. (2025); Lee et al. (2025), regardless of whether candidates are tightly clustered or widely scattered, leaving the crop either too broad to improve resolution or too narrow to retain critical context.
1
(a) Direct Grounding (UI-R1, GUI-G2)
(b) Iterative Cropping (DiMo-GUI, Nguyen, RegionFocus)
...
Crop × N Dense-interface Limitation
Predict
Large Resource Costs
(c) UI-Zoomer (Ours)
Rigid Cropping Ratio
Choice I: Consensus Voting Reliability
TTS Stronger Robustness
Gating One-take Time Costs
Choice II: Adaptive Cropping
Predict
Figure 1 Comparison of GUI grounding paradigms. (a) Direct grounding methods struggle with dense interfaces. (b) Iterative cropping methods incur large resource costs and use rigid cropping ratios. (c) Our UI-Zoomer applies Test-Time Scaling (TTS) with reliability gating, adaptively choosing between consensus voting and adaptive cropping, achieving stronger robustness with one-take time costs.
The root cause is that these methods treat all instances Method Avg Acc Time uniformly, without consulting the model’s own prediction w/o DiMo-GUI 81.84% 35:47 behavior. Recent work shows that spatial agreement across w/ DiMo-GUI 77.20% 6:43:07 stochastic samples correlates with localization reliability Du et al. (2025), and that coordinate likelihoods near a preTable 1 Accuracy and inference time of w/ and dicted point follow a smooth Gaussian distribution in pixel w/o iterative cropping on ScreenSpot-V2. space Lee et al. (2025), confirming that VLMs implicitly encode continuous spatial uncertainty. The variance of sampled predictions Wang et al. (2026); Du et al. (2025) thus encodes both whether the model is confused and over what spatial extent, which is precisely the information needed to gate zoom-in and size the crop window. This motivates a simple but previously unexplored principle: zoom only when uncertain, and zoom by how much the predictions disagree. Building on this insights, we propose UI-Zoomer, a training-free adaptive zoom-in framework for GUI grounding. UI-Zoomer first draws N stochastic candidates from the model and computes a reliability score by fusing spatial consensus with token-level confidence; instances that pass the gate are resolved immediately by consensus voting. For uncertain instances, the crop window is derived from the variance of candidate predictions decomposed into inter-sample positional spread and intra-sample box extent, yielding a per-instance radius that contracts for easy cases and expands for hard ones. A single deterministic re-inference pass on the resulting crop completes the refinement. Extensive experiments on three widely-adopted GUI grounding benchmarks, ScreenSpot-Pro, UI-Vision, and ScreenSpot-v2, demonstrate that UI-Zoomer consistently improves over strong baselines, achieving gains of up to +13.4%, +10.3%, and +4.2% respectively. Icon targets benefit more than text targets on average, consistent with the intuition that compact and semantically ambiguous elements profit most from high-resolution refinement. Ablations confirm the independent contribution of each component and the advantage of adaptive crop sizing over any fixed-ratio alternative. Our contributions are threefold: • We propose UI-Zoomer, a training-free adaptive zoom-in framework that frames the trigger and scale of zoom-in as a prediction uncertainty quantification problem. • UI-Zoomer comprises a confidence-aware gate that avoids unnecessary computation by routing only
2
uncertain instances to refinement, and a Gaussian-based adaptive crop sizing module that derives the crop window from the variance of candidate predictions. • Extensive experiments on ScreenSpot-Pro, UI-Vision, and ScreenSpot-v2 demonstrate consistent improvements across four model architectures, with gains of up to +13.4% on ScreenSpot-Pro.
2
Related Work
2.1
GUI Grounding
GUI grounding requires predicting the pixel coordinates of an interface element given a screenshot and a natural language instruction. Early work builds pipeline-based systems that chain OCR, icon detectors, and LLMs for planning and element selection Zhang et al. (2025); Wang et al. (2024); Li et al. (2024); Agashe et al. (2024); Zhang et al. (2025); Liu et al. (2024); Yang et al. (2023); Bai et al. (2021); Xu et al. (2024); Wu et al. (2024); Tang et al. (2025b); Wu et al. (2025b); Agashe et al. (2025). A second generation trains specialist VLMs end-to-end on large-scale GUI corpora, with models such as UGround Gou et al. (2024), OS-Atlas Wu et al. (2024), and UI-TARS Qin et al. (2025) demonstrating strong cross-platform generalization. More recently, reinforcement fine-tuning has emerged as a data-efficient alternative: methods including UI-R1 Lu et al. (2025a), GUI-G2 Tang et al. (2025a), SE-GUI Yuan et al. (2025), and UI-Venus Gu et al. (2025) apply GRPO-style objectives with coordinate accuracy rewards, matching or exceeding SFT models trained on orders of magnitude more data. Despite these advances, all training-time approaches share a hard ceiling at high resolution: once a target element is too small to resolve in a standard forward pass, additional training provides diminishing returns Wu et al. (2025a); Luo et al. (2025).
2.2
Test-Time Scaling for GUI Grounding
Test-time scaling improves model performance at inference without modifying parameters Snell et al. (2024). In GUI grounding, the dominant paradigm is zoom-in inference: DiMo-GUI Wu et al. (2025a) applies iterative zoom-in with a fixed crop ratio; RegionFocus Luo et al. (2025) triggers zoom-in upon execution errors; ReGUIDE Lee et al. (2025) uses KDE over multiple predictions to identify a high-density crop center; Nguyen Nguyen (2024) proposes successive iterative narrowing. A parallel thread exploits prediction consistency as a reliability signal: GUI-RC Du et al. (2025) constructs spatial voting grids over stochastic samples to identify consensus regions; SafeGround Wang et al. (2026) derives calibrated uncertainty estimates from spatial dispersion with statistical guarantees; GUI-Eyes Chen et al. (2026) trains models via RL to actively decide when to invoke zoom tools. These methods either apply cropping regardless of per-instance confidence, or use consistency signals purely for voting without connecting them to crop sizing. UI-Zoomer unifies both perspectives by using prediction variance to simultaneously gate zoom-in and derive per-instance crop windows.
3
Method
3.1
Problem Setup
Given a GUI screenshot I ∈ RH×W ×3 and a natural-language instruction q, we predict a click location p̂ ∈ [0, 1]2 in normalized image coordinates. We represent each localization hypothesis as an axis-aligned bounding box b = [x1 , y1 , x2 , y2 ] and define the click as its center: x1 + x2 y1 + y2 p̂ = , . (1) 2 2 As shown in Figure 2, UI-Zoomer proceeds in three stages: (1) global multi-sampling, (2) reliability gating, and (3) adaptive crop and zoom. The full procedure is summarized in Algorithm 1.
3
Open Spell Check for US English.
(a) Test Time Scaling (TTS)
(b) Reliability Gating
(c) Choice I: Voting
Consistency
Re Mor lia e bl e
Re Les lia s bl e
Con�idence
Choice I: Voting
Rule 1: Pairwise IoU Rule 2: Con�idence
Choice II: Adaptive Cropping
SS-v2 Accuracy
(d) Choice II: Adaptive Cropping
Cropping
2D Gaussian Modeling
Variance Decomposition
Final Prediction
UI-Zoomer (Ours)
Dimo-GUI Increasing Grounding Ef�iciency Time Cost
Figure 2 Overview of UI-Zoomer. (a) The model samples N candidate predictions via Test-Time Scaling (TTS). (b) A reliability gate routes confident instances to consensus voting (Choice I) and uncertain ones to adaptive cropping (Choice II). (d) The crop window is derived from 2D Gaussian variance decomposition, enabling per-instance adaptive zoom-in.
Algorithm 1 UI-Zoomer Require: Image I, instruction q, model M, N , threshold τ , scale γ, min crop m Ensure: Click point p̂ ∈ [0, 1]2 1: {bi , ci }N Stage 1: global multi-sampling i=1 ← Sample(M, I, q; T =0.9) 2: Compute Cspatial (Eq. 3), c̄ (Eq. 2), S = Cspatial + c̄ Stage 2: reliability gating 3: if S > τ then 4: return center(bi⋆ ) pass: consensus vote (Eq. 5) 5: else 6: µ , σ ← FilterAndDecompose({bi }) Stage 3: filter + variance decomp. 7: (xc1 , y1c , xc2 , y2c ) ← AdaptiveCrop( µ , σ ; γ, m) adaptive crop window (Eq. 10) 8: b̂ ← M(Crop(I, xc1 , y1c , xc2 , y2c ); T =0) 9: return center(MapBack(b̂)) 10: end if
3.2
zoom: deterministic re-inference map back to global coords (Eq. 11)
Stage 1: Global Multi-Sampling
We sample N =8 candidate boxes from M at temperature T =0.9 and discard invalid parses. For each valid candidate i we record the predicted box bi and estimate a scalar confidence from the geometric mean of token probabilities: ! Li 1 X ci = exp log pi,t , (2) Li t=1 where Li is the sequence length and pi,t is the probability of the t-th token.
3.3
Stage 2: Reliability Gating
When candidates are consistent and confident, zoom-in is unnecessary and costly. We quantify this reliability through two complementary signals and use their combination to selectively trigger refinement.
4
3.3.1
Spatial consensus.
We quantify cross-sample agreement by the mean pairwise IoU: X 1 Cspatial = IoU(bi , bj ). N (N − 1)
(3)
i̸=j
3.3.2
Gating score.
We combine spatial consensus with average token confidence: N
S = Cspatial + c̄,
c̄ =
1 X ci . N i=1
(4)
The two signals are complementary: Cspatial is sensitive to positional scatter while c̄ reflects sharpness of the predictive distribution over coordinate tokens. When S > τ , we trust the global predictions and return immediately. 3.3.3
Consensus voting.
We select the candidate with the most peer support, breaking ties by confidence: X vi = I[IoU(bi , bj ) > 0.5], i⋆ = arg max (vi , ci ). i
j̸=i
3.4
(5)
Stage 3: Uncertainty-Driven Adaptive Crop
When S ≤ τ , candidates are unreliable and zoom-in is warranted. Rather than using a fixed crop ratio, we derive the crop window directly from the variance of the candidate set. 3.4.1
Outlier filtering.
A small number of erratic samples can inflate the estimated variance and produce an oversized crop. We therefore discard outliers by retaining only the K = ⌊0.75N ⌋ candidates whose centers lie closest to the median center z̃: di = ∥zi − z̃∥2 , K = arg topK {−di }, (6) i
where zi denotes the center of bi . We compute subsequent statistics over K. 3.4.2
Variance decomposition.
We model the unknown target location Z as a latent random variable and apply the law of total variance coordinate-wise: Var(Z) = Var(E[Z | I]) + E[Var(Z | I)] . (7) | {z } | {z } vinter
vintra
The inter-sample term captures positional disagreement across draws: 1 X 1 X (zi − µ)⊙2 , µ= zi . vinter = K K i∈K
(8)
i∈K
The intra-sample term encodes the predicted scale of each element. Treating each box as a Gaussian spanning ±2σ of its width and height: 1 X si ⊙2 vintra = , (9) K 4 i∈K
⊤
where si = [six , siy ] is the width and height of bi . The two terms are complementary: vinter expands the crop when candidates disagree on position; vintra ensures the crop is at least as large as the predicted element even when candidates coincide. 5
3.4.3
Crop window.
√ We set the crop radius as r = γσ, where σ = vinter + vintra . To avoid degenerate crops and aspect-ratio distortions, we impose a minimum side length m and squarify: s = max(2rx , 2ry , m),
[xc1 , y1c , xc2 , y2c ] = [µx − 2s , µy − 2s , µx + 2s , µy + 2s ].
(10)
If the window extends beyond image boundaries, we shift it inward while preserving its size. 3.4.4
Zoom and map back.
We crop I to this window, resize it to the model’s resolution budget, and run a single deterministic pass (T =0) to obtain a refined box b̂ in crop coordinates. We map it back to global normalized coordinates via: x=
xc1 + x̂ wc , W
y=
y1c + ŷ hc , H
(11)
where wc = xc2 − xc1 and hc = y2c − y1c . If refinement produces an invalid box, we fall back to the most confident global candidate.
4
Experiments
4.1
Setup
4.1.1
Benchmarks.
We evaluate on three benchmarks spanning different difficulty regimes. ScreenSpot-Pro Li et al. (2025) targets 4K professional desktop environments across 23 applications, with unusually small and dense targets. ScreenSpot-v2 Wu et al. (2024) is a multi-platform benchmark covering mobile, desktop, and web interfaces with 1,200+ instructions. UI-Vision Nayak et al. (2025) covers fine-grained desktop grounding across 83 real-world applications, including element grounding, layout grounding, and action prediction. Following prior work Cheng et al. (2024); Li et al. (2025), we report click accuracy: a prediction is correct if the output point falls within the ground-truth bounding box. 4.1.2
Models.
We evaluate our method using two categories of base models: (1) general-purpose VLMs, i.e., Qwen2.5-VL7B Bai et al. (2025), an open-source multimodal model pretrained on large-scale data; and (2) GUI-specific VLMs, including UI-Venus-7B, UI-Venus-72B Gu et al. (2025), and GUI-G2 -7B Tang et al. (2025a), which are tailored for GUI understanding and grounding. Notably, both UI-Venus and GUI-G2 are further enhanced with reinforcement learning, leading to stronger task-specific alignment for UI interaction and more reliable GUI grounding behaviors. 4.1.3
Implementation.
All evaluations are conducted on 4 NVIDIA RTX 4090D 24G GPUs. We use the vLLM engine with a context length of 16,384 tokens. We sample N =8 candidates at temperature T =0.9 and set the minimum crop side to m=512 pixels. The gating threshold τ and Gaussian scale γ are tuned per model-benchmark pair; for UI-Venus-7B on ScreenSpot-Pro we use τ =1.0 and γ=2.5.
4.2
Main Results
Table 3 reports results on ScreenSpot-Pro; UI-Vision and ScreenSpot-v2 results appear in Table 2 (Full results are provided in Appendix Tables 11 and 10). UI-Zoomer consistently improves all four models across all three benchmarks, with average gains of up to +13.4%, +10.3%, and +4.2% on ScreenSpot-Pro, UI-Vision, and ScreenSpot-v2 respectively.
6
Category 1 text icon avg
text
UI-Venus-7B + UI-Zoomer ∆ Improvement
99.0 98.6 -0.4
90.1 90.5 +0.4
95.2 95.2 +0.0
97.9 99.0 +1.1
89.3 92.9 +3.6
UI-Venus-7B + UI-Zoomer ∆ Improvement
68.5 80.5 +12.0
23.8 32.3 +8.5
33.3 42.5 +9.2
65.5 74.7 +9.2
19.9 30.6 +10.7
Benchmark
Methods
ScreenSpot-v2
UI-Vision
Category 2 icon avg
Category 3 text icon avg
text
Overall icon
avg
94.3 96.4 +2.1
94.9 95.7 +0.8
89.7 90.6 +0.9
92.5 93.4 +0.9
97.4 97.8 +0.4
89.7 91.2 +1.5
94.0 94.9 +0.9
29.7 40.1 +10.4
38.7 55.9 +17.2
7.8 15.1 +7.3
11.4 19.7 +8.3
60.6 72.7 +12.1
16.5 25.2 +8.7
24.4 33.7 +9.3
Table 2 Performance of UI-Venus-7B with and without UI-Zoomer on ScreenSpot-v2 (Mobile / Desktop / Web) and UI-Vision (Basic / Functional / Spatial). Development text icon
Creative text icon
CAD text icon
Scientific text icon
Office text icon
text
icon
text
Overall icon
avg
Proprietary Methods GPT-4o Hurst et al. (2024) Claude-3.7-Sonnet cla Seed-1.5-VL Guo et al. (2025)
1.3 -
0.0 -
1.0 -
0.0 -
2.0 -
0.0 -
2.1 -
0.0 -
1.1 -
0.0 -
0.0 -
0.0 -
1.3 -
0.0 -
0.8 27.7 60.9
General Open-source Models OS-Atlas-7B Wu et al. (2024) Qwen2.5-VL-3B Bai et al. (2025) UGround-7B Gou et al. (2024) UGround-72B UI-TARS-7B UI-TARS-72B Jedi-7B Xie et al. (2025) Qwen2.5-VL-32B
33.1 38.3 58.4 63.0 42.9 74.0
1.4 3.4 12.4 17.3 11.0 21.4
28.8 40.9 50.0 57.1 50.0 61.1
2.8 4.9 9.1 15.4 11.9 13.3
12.2 22.3 20.8 18.8 38.0 38.1
4.7 6.3 9.4 12.5 14.1 15.6
37.5 44.4 63.9 64.6 72.9 78.5
7.3 10.0 31.8 20.9 25.5 29.1
33.9 48.0 63.3 63.3 75.1 76.3
5.7 17.0 20.8 26.4 47.2 37.7
27.1 33.6 30.8 42.1 33.6 55.1
4.5 4.5 16.9 15.7 16.9 27.0
28.1 37.8 47.8 50.9 52.6 63.2
4.0 6.6 16.2 17.5 18.2 22.5
18.9 25.9 31.1 34.5 35.7 38.1 39.5 47.6
Reinforcement Learning Methods UI-TARS-1.5 Qin et al. (2025) GTA1-7B Yang et al. (2025) UI-R1-E-3B Lu et al. (2025a) UI-S1-7B Lu et al. (2025b) SE-GUI-7B Yuan et al. (2025)
53.3 46.1 51.3
17.2 6.9 42.2
66.9 41.9 68.2
20.7 4.2 19.3
62.6 37.1 57.6
18.2 12.5 9.1
76.4 56.9 75.0
31.8 21.8 28.2
82.5 65.0 78.5
50.9 26.4 43.4
48.6 32.7 49.5
25.9 10.1 25.8
65.5 63.5
25.2 21.0
61.6 50.1 33.5 30.6 47.3
Test Scaling Methods DiMo-GUI Wu et al. (2025a) RegionFocus Luo et al. (2025) GUI-RC Du et al. (2025) UI-Venus-7B [pass@4] UI-Venus-7B [pass@8]
66.9 53.2 77.9 81.2
21.4 3.4 29.0 32.4
60.6 42.9 68.0 70.1
21.7 4.9 19.6 24.5
50.3 28.4 66.0 69.5
14.1 3.1 25.00 28.1
68.1 56.9 79.2 81.3
21.8 10.9 26.4 29.1
80.8 59.9 83.2 87.0
52.8 24.5 37.7 43.4
69.2 41.1 58.9 66.4
28.1 15.7 25.8 27.0
65.2 46.6 72.6 75.8
24.5 8.8 26.2 29.6
49.7 32.1 24.0 54.8 58.2
Our method Qwen2.5-VL-7B + UI-Zoomer ∆ Improvement
48.7 63.6 +14.9
2.1 17.9 +15.8
32.0 45.7 +13.7
4.9 14.0 +9.1
24.4 51.3 +26.9
4.7 14.1 +9.4
51.4 47.2 -4.2
7.3 20.0 +12.7
53.7 66.3 +12.6
18.9 34.0 +15.1
38.3 49.5 +11.2
10.1 28.1 +18.0
40.6 54.0 +13.4
6.6 19.9 +13.3
27.6 41.0 +13.4
GUI-G2 -7B Tang et al. (2025a) + UI-Zoomer ∆ Improvement
67.5 79.9 +12.3
24.1 38.6 +14.5
59.9 68.0 +8.1
16.1 26.6 +10.5
55.3 77.7 +22.3
20.3 34.4 +14.1
75.7 82.6 +7.0
28.2 36.4 +8.2
75.8 84.3 +8.4
39.6 60.4 +20.8
50.5 65.4 +15.0
20.2 38.2 +18.0
64.4 76.7 +12.3
23.3 36.8 +13.4
48.7 61.4 +12.7
UI-Venus-7B Gu et al. (2025) + UI-Zoomer ∆ Improvement
72.7 80.5 +7.8
22.8 37.2 +14.4
62.4 70.1 +7.7
15.4 31.5 +16.1
58.9 77.2 +18.3
21.9 34.4 +12.5
74.3 82.6 +8.3
26.4 30.0 +3.6
78.7 88.8 +10.1
35.9 50.9 +15.0
50.5 67.3 +16.8
23.6 37.1 +13.5
66.7 78.1 +11.4
22.9 35.4 +12.5
50.0 61.8 +11.8
UI-Venus-72B + UI-Zoomer ∆ Improvement
80.5 85.7 +5.2
32.4 42.1 +9.7
70.1 75.1 +5.0
32.9 44.8 +11.9
63.5 76.1 +12.6
29.7 40.6 +10.9
75.0 84.0 +9.0
39.1 42.7 +3.6
83.7 86.5 +2.8
49.1 69.8 +20.7
73.8 83.2 +9.4
34.8 48.3 +13.5
74.0 81.3 +7.3
35.3 46.0 +10.7
59.2 67.8 +8.6
Methods
OS
Table 3 Performance comparison on ScreenSpot-Pro across four models: Qwen2.5-VL-7B, GUI-G2 -7B, UI-Venus-7B, and UI-Venus-72B. For a fair comparison, RegionFocus is evaluated using Qwen2.5-VL-7B as the backbone.
Zoom-in is most effective where resolution matters most. Gains are largest on ScreenSpot-Pro, the highest-resolution benchmark, and smallest on ScreenSpot-v2, which covers standard-resolution mobile and web interfaces. Within ScreenSpot-Pro, icon targets benefit more than text targets across all models (+12.5% vs. +11.1%), consistent with the intuition that compact and semantically ambiguous elements are most limited by resolution in a single forward pass. Adaptive zoom outperforms both naive sampling and prior test-time methods. Compared to naive sampling baselines (UI-Venus-7B pass@4: 54.84%, pass@8: 58.19%), UI-Zoomer reaches 61.8% at a comparable inference budget. It also substantially outperforms the prior zoom-in method RegionFocus Luo et al. (2025) (32.1%), which applies cropping unconditionally with a fixed ratio. Against RL-trained methods, UI-Zoomer with UI-Venus-7B surpasses UI-S1-7B (30.6%) by +31.2% and GTA1-7B (50.1%) by +11.7% on ScreenSpot-Pro, 7