Pure-vision GUI agents provide universal interaction capabilities but suffer from severe efficiency bottlenecks due to the massive spatiotemporal redundancy inherent in high-resolution screenshots and historical trajectories. We identify two critical misalignments in existing compression paradigms: the temporal mismatch, where uniform history encoding diverges from the agent's "fading memory" attention pattern, and the spatial topology conflict, where unstructured pruning compromises the grid integrity required for precise coordinate grounding, inducing spatial hallucinations. To address these challenges, we introduce GUIPruner, a training-free framework tailored for high-resolution GUI navigation. It synergizes Temporal-Adaptive Resolution (TAR), which eliminates historical redundancy via decay-based resizing, and Stratified Structure-aware Pruning (SSP), which prioritizes interactive foregrounds and semantic anchors while safeguarding global layout. Extensive evaluations across diverse benchmarks demonstrate that GUIPruner consistently achieves state-of-the-art performance, effectively preventing the collapse observed in large-scale models under high compression. Notably, on Qwen2-VL-2B, our method delivers a 3.4x reduction in FLOPs and a 3.3x speedup in vision encoding latency while retaining over 94% of the original performance, enabling real-time, high-precision navigation with minimal resource consumption.
Figures & tables
Figure 1. Paradigm comparison of visual encoding and pruning. In contrast to conventional pipelines (Top) that incur high redundancy via uniform encoding and disrupt topology via unstructured pruning, GUIPruner (Bottom) implements efficient source-level reduction and structured, topology-preserving compression, ensuring high-precision grounding with minimal token consumption.
Figure 2. (a) Cross-attention weights exhibit a pronounced Temporal Decay , consistent with the “Recency Effect” in working memory. (b) Background tokens dominate the visual context across four datasets, indicating significant spatial redundancy. (c) Heatmaps reveal that specific background regions (red boxes) retain high attention as essential semantic anchors, cautioning against indiscriminate pruning.
Figure 3. Overview of the GUIPruner framework. The framework addresses spatiotemporal redundancy through two synergistic modules: (Left) Temporal-Adaptive Resolution (TAR) mimics biological fading memory by assigning decaying resolution budgets to historical frames based on temporal distance, eliminating redundancy in distant context. (Right) Stratified Structure-aware Pruning (SSP) operates on the current frame within shallow LLM layers. It preserves topological integrity by hierarchically retaining interactive foreground tokens ( Sfg ), semantic background anchors ( Sbg ), and a uniform structural grid ( Suni ), effectively compressing visual tokens without inducing spatial hallucinations.
Method
Qwen2-VL-2B
Qwen2.5-VL-7B
AITW
M2Web
GUI-Ody
AndCtrl
AITW
M2Web
GUI-Ody
AndCtrl
Upper Bound (Full Tokens)
Original Model
69.5
34.5
71.8
66.9
71.2
35.2
78.4
71.6
Setting I: Retain History 40% / Current 75%
FastV (ECCV’24)
66.2
33.4
67.3
66.3
67.9
34.3
73.5
70.8
DivPrune (CVPR’25)
65.3
22.8
67.1
65.6
23.2
7.7
13.6
29.4
Table 1. Performance benchmarking on GUI navigation tasks. We evaluate Qwen2-VL-2B and Qwen2.5-VL-7B across three pruning settings. Abbreviations: “GUI-Ody” (GUI-Odyssey), “M2Web” (Mind2Web), and “AndCtrl” (AndroidControl).
Method
# Token
FLOPs (T)
Encoder Time (ms)
Prefill Time (ms)
GPU Memory (MB)
Qwen2-VL-2B
1320
11.5
87.9
47.5
8956
FastV
310
8.7 ( × 1.3)
87.9 ( × 1.0)
28.6 ( × 1.6)
6798
DivPrune
310
8.5 ( × 1.4)
87.9 ( × 1.0)
27.5 ( × 1.7)
6180
CDPruner
310
8.5 ( × 1.4)
87.9 ( × 1.0)
27.4 ( × 1.7)
6180
MoB
310
8.7 ( × 1.3)
87.9 ( × 1.0)
28.1 ( × 1.6)
6734
GUIPruner
310
3.4 ( × 3.4)
26.6 ( × 3.3)
24.1 ( × 1.9)
5902
Table 2. Efficiency comparison on Qwen2-VL-2B (AITW dataset, 4 history frames, RTX 4090). We report the number of tokens, FLOPs, Encoder latency, Prefill latency, and GPU memory usage. Bold indicates the best performance.
Figure 4. Decoupled sensitivity analysis of TAR and SSP on Qwen2-VL-2B. Left: Evaluation of TAR under varying history retention ratios ( τ ) on AITW and Mind2Web. Right: Evaluation of SSP under varying current frame retention ratios ( κ ). GUIPruner consistently demonstrates superior robustness over baselines, particularly in high-compression regimes.
Method
AITW
GUI-Ody
Avg.
Rel.
Qwen2-VL-2B
Upper Bound (Full Tokens)
Vanilla
69.5
71.8
70.6
100%
Qwen2-VL-2B
Retain History 40% / Current 100%
TAR w/ Uniform
68.4
70.2
69.3
98.2%
TAR (Ours)
69.4
70.6
70.0
99.2%
Qwen2-VL-2B
Retain History 20% / Current 100%
Table 3. History token allocation ablation. We compare our proposed decay-based strategy (TAR) against a uniform allocation baseline across varying history budgets. "Rel." denotes performance relative to the full-context upper bound.
Method
AITW
GUI-Ody
Avg.
Rel.
Qwen2-VL-2B
Upper Bound (Full Tokens)
Vanilla
69.5
71.8
70.6
100%
Qwen2-VL-2B
Retain History 100% / Current 75%
SSP w/ Random
66.5
69.4
68.0
96.3%
SSP (Ours)
67.7
69.8
68.8
97.4%
Qwen2-VL-2B
Retain History 100% / Current 60%
Table 4. Comparison of spatial sampling strategies for current frame pruning. We assess SSP with uniform grid sampling versus random sampling.
Figure 5. Hyperparameter sensitivity analysis on AITW. We evaluate the Step SR under varying compression intensities: the impact of the temporal decay factor γ in TAR, and the impact of the background saliency factor ρ in SSP.
L
1
2
4
8
16
Step SR
12.4
65.3
64.4
61.1
60.6
FLOPs (T)
3.79
3.39
3.45
3.48
3.53
Table 5. Parameter study on the pruning layer depth ( L ) for SSP. We analyze the trade-off between navigation performance (Step SR) and computational cost (FLOPs) by varying the layer index where pruning is executed.
Appendix figures & tables6 assets
Supplementary material from the paper’s appendix.
Appendix
Table 11
Method
# Token
FLOPs (T)
Encoder Time (ms)
Prefill Time (ms)
GPU Memory (MB)
Qwen2.5-VL-7B
1320
27.5
308.7
156.3
18004
FastV
310
13.8 ( × 2.0)
308.7 ( × 1.0)
59.1 ( × 2.6)
17680
DivPrune
310
12.7 ( × 2.2)
308.7 ( × 1.0)
53.6 ( × 2.9)
16616
CDPruner
310
12.7 ( × 2.2)
308.7 ( × 1.0)
53.8 ( × 2.9)
16614
MoB
310
13.8 ( × 2.0)
308.7 ( × 1.0)
58.4 ( × 2.7)
16768
GUIPruner
310
8.9 ( × 3.1)
97.7 ( × 3.2)
52.8 ( × 3.0)
16386
Appendix
Table 6. Efficiency comparison on Qwen2.5-VL-7B (AITW dataset, 4 history frames, RTX 4090). We report the number of tokens, FLOPs, Encoder latency, Prefill latency, and GPU memory usage. Bold indicates the best performance.
Figure 6. Visualizing the SSP Process. A step-by-step breakdown of token retention. The method progressively layers Foreground (Red), Saliency-based Background (Green), and Uniform Grid (Blue) tokens. This guarantees that both fine-grained interactive elements and the coarse-grained global layout are preserved for the VLM.
Figure 7. Evolution of LLM Attention Dynamics across Pruning Depths. The heatmaps illustrate the LLM’s visual attention for search-related queries. (L1) Attention is misaligned, failing to cover the target search bar. (L2) Attention is precisely focused on the search bar , providing the most accurate signal for token pruning. (L4–L16) Attention progressively expands to the global context, reducing the selectivity required for efficient pruning.
Figure 8. Visualization on AITW (Mobile GUI). Qualitative comparison demonstrating that baselines produce fragmented, spatially disjoint patterns. In contrast, GUIPruner preserves the global layout integrity of history frames and enforces a uniform, structure-aware topology for the current frame.
Figure 9. Visualization on Mind2Web (Web GUI). Qualitative comparison demonstrating that GUIPruner effectively preserves the global layout of historical webpages while precisely highlighting interactive elements in the current frame, mitigating the spatial fragmentation observed in baselines.