cs.CVFeb 27, 2026

ST-Lite: Training-Free KV Cache Compression with Spatio-Trajectory Guidance for Long-Horizon GUI Agents

Authors: Bowen Zhou, Zhou Xu, Wanli Li, Jingyu Xiao, Pingan Gan, Haoqian Wang

Organizations: Tsinghua University, Shenzhen, China · Zhejiang University, Hangzhou, China · The Chinese University of Hong Kong, Hong Kong, China

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Training-free KV cache compression is essential for deploying vision-language GUI agents under memory and latency constraints, yet existing methods are designed for generic language workloads and ignore the distinctive structure of GUI interaction traces. We characterize three GUI-specific workload properties--high inter-frame visual redundancy, extremely small UI-element spatial footprints, and near-uniform cross-layer attention sparsity--that cause existing schemes to retain as few as 39% of oracle-important KV pairs at the standard 20% budget. To address this, we propose ST-Lite, a training-free compression scheme whose three components each target one property: Trajectory-aware Semantic Gating (TSG) filters redundant historical frames, Component-centric Spatial Saliency (CSS) preserves fine-grained element boundaries, and a flat per-layer budget avoids hierarchical misallocation. Across seven GUI benchmarks and two backbones in the deployment-relevant 10%-40% window, ST-Lite consistently outperforms all existing compression baselines, matching or exceeding Full Cache task accuracy on the primary backbone at the 20% budget while delivering up to 2.35x decoding speedup at fivefold compression. The implementation is available at https://github.com/94wen94/ST-Lite.

Figures & tables

Figure 1: Conceptual illustration of ST-Lite compared with existing methods. The left side depicts the limitations of current strategies: window-based greedy methods (e.g., SnapKV) suffer from local optimum traps , while hierarchical allocation methods (e.g., PyramidKV) lead to structure misalignment . The right side demonstrates our ST-Lite, which integrates Component-centric Spatial Saliency (CSS) to preserve key interactive elements and Trajectory-aware Semantic Gating (TSG) to eliminate history redundancy, achieving efficient and accurate long-horizon reasoning.
Figure 2: Task execution flows under different context settings. (a) The original sequence O\mathcal{O} with full historical frames leads to task success. (b) The pruned sequence Or\mathcal{O}_{r} , with trajectory-wise redundant frames removed, also succeeds, confirming the high historical redundancy in GUI trajectories.
Figure 3: Layer-wise attention sparsity. GUI agents on (a) AITW and (b) AgentNetBench exhibit near-uniform high sparsity across all transformer layers, in contrast to the pyramid-shaped profiles measured on text and general vision-language tasks.
Figure 4: The overall architecture of ST-Lite. Our scheme dynamically optimizes the KV cache through two complementary modules: (1) Component-centric Spatial Saliency (CSS) , which identifies and preserves spatially salient regions (e.g., functional buttons) within each frame using local hidden-state cosine similarity within a 3×33{\times}3 neighborhood; and (2) Trajectory-aware Semantic Gating (TSG) , which filters out redundant historical states by measuring semantic similarity between historical tokens and the current frame. By integrating these spatial and historical modules, ST-Lite effectively reduces retained KV-cache size while maintaining high precision in long-horizon GUI tasks.
Figure 5: Evaluation results on ScreenSpot Pro, AITW, and AgentNetBench with varied cache budgets. ST-Lite achieves comparable accuracy against Full Cache and outperforms multiple baselines with limited KV cache budget. On AITW, ST-Lite exceeds Full Cache at β=20%\beta{=}20\% ; this less-is-more effect is analysed with paired statistical tests in § V .
Figure 6: Cache Hit Rate on AITW (UI-TARS-1.5-7B) across β∈{3,5,10,15,20}%\beta\in\{3,5,10,15,20\}\% . ST-Lite leads SnapKV, PyramidKV and VL-Cache at every budget, with a +13.0%+13.0\% gap to SnapKV at β=20%\beta=20\% , peaking through β∈{10,15}%\beta\in\{10,15\}\% (smaller at the budget extremes).
MethodScreenSpot ProAITWAgentNet
Full Cache ( β=100%\beta{=}100\% )42.318.217.5
SnapKV ( β=20%\beta{=}20\% )39.715.39.3
ST-Lite (CSS only)42.217.414.1
ST-Lite (TSG only)39.618.516.1
ST-Lite (CSS+TSG)42.220.117.0
Table 1: Ablation contributions of CSS and TSG at β=20%\beta{=}20\% . All rows use the same default configuration (SnapKV-style window base, δ=8\delta{=}8 ). CSS-only sets α=1\alpha{=}1 , TSG off; TSG-only sets α=0\alpha{=}0 , TSG on.
Figure 7: Impact of historical frame count on model success rate. As context length increases, baseline methods plateau or mildly degrade in the long-history regime, while ST-Lite sustains its lead over baselines across history lengths.
ScreenshotsPrefill SpeedupDecoding SpeedupEnd-to-End Speedup
30.971.551.30
50.981.951.55
100.992.351.80
Table 2: Inference speedup across screenshot scales at β=20%\beta{=}20\% . The prefill phase incurs at most 11 – 2%2\% overhead; ST-Lite accelerates the memory-bound decoding phase up to 2.35 ×\times .
Appendix figures & tables43 assets

Supplementary material from the paper’s appendix.

Appendix

β\betaST-LiteSnapKVPyrKVGUI-KVVL-Cache
3%3\%38.728.315.829.512.0
5%5\%47.931.526.134.421.5
10%10\%58.642.037.543.732.5
15%15\%64.347.741.752.738.0
20%20\%66.153.140.454.139.0
Appendix
Table 3: Cache Hit Rate (%) on AITW at five low-budget operating points. ST-Lite leads at every budget; gap to the strongest baseline peaks at +14.9%+14.9\% at β=10%\beta{=}10\% (ST-Lite 58.658.6 , GUI-KV 43.743.7 ).
β\betaST-LiteSnapKVPyrKVGUI-KVVL-Cache
5%5\%41.327.122.030.718.8
10%10\%54.138.533.040.828.4
20%20\%62.849.538.151.436.2
Appendix
Table 4: CHR (%) on AgentNetBench (UI-TARS-1.5-7B). ST-Lite leads at every budget, with a +11.4%+11.4\% gap over the strongest baseline (GUI-KV) at β=20%\beta{=}20\% . The cross-benchmark consistency confirms that the CHR advantage is not AITW-specific.
Figure 8: A representative consecutive frame pair from AgentNetBench (task s_02e17403c227ac27 ). (a) frame tt . (b) frame t+1t{+}1 . (c) pixels where the two frames disagree by more than 0.01 in any RGB channel, highlighted in red on top of frame t+1t{+}1 : only 13.6%13.6\% of pixels change between the two frames. The other 86.4%86.4\% are essentially identical. This is in the typical range for AgentNetBench (median 82.7%82.7\% unchanged).
Statisticp10p25p50p75p90
Mean abs. pixel diff (0-1)0.0010.0040.0180.0790.229
% pixels essentially unchanged18.352.282.795.398.6
Appendix
Table 5: Inter-frame redundancy statistics on AgentNetBench (311 consecutive frame pairs, 30 tasks). The median frame pair shares 82.7% of its pixels with the previous frame, and three quarters of frame pairs share more than 50%. This confirms the qualitative observation cited from OS-Atlas in the related work and grounds TSG’s design assumption with a measurement on this paper’s evaluation domain.
Statisticp25p50p75p90
Width (% screen)0.901.674.457.54
Height (% screen)1.201.562.063.17
Area (% screen)0.010.040.080.13
Width (pixels)2857140238
Height (pixels)21253358
Appendix
Table 6: UI element size statistics across all 1,581 ScreenSpot-Pro annotations. Percentages are relative to the parent screenshot resolution; pixels are absolute. The median UI element occupies roughly 1.7%×1.6%1.7\%\times 1.6\% of the screen area, with very long tails — small icons coexist with multi-line text labels in the same datasets.
StatisticAITWAgentNetBench
Mean sparsity97.37%97.66%
Std. dev. across layers1.26%1.29%
Coefficient of variation (CV)1.30%1.32%
Range (max −- min layer)4.98%4.81%
Largest adjacent-layer gradient3.79%3.31%
Appendix
Table 7: Cross-layer sparsity dispersion on OpenCUA-7B (28 layers, 20 trajectories per benchmark). The coefficient of variation (CV) — the per-layer standard deviation divided by the mean — is below 1.4%1.4\% on both benchmarks. Equivalently, every layer’s sparsity falls within a ≤5%\leq 5\% range, and consecutive layers never differ by more than 3.8%3.8\% . This is the operational definition of “uniform” adopted in § 2 .
LayerAITW (%)AgentNetBench (%)
095.394.5
498.898.9
897.898.1
1296.797.0
1696.897.5
2098.698.8
Appendix
Table 8: Per-layer attention sparsity (%) of OpenCUA-7B on AITW and AgentNetBench. Sparsity remains ≥94.5%\geq 94.5\% at every probed layer with negligible cross-layer variation, consistent with the uniform-allocation justification in § 2 .
Base scoringBase+CSS+TSGFull
Mean Attention16.4718.1217.8918.19
Uniform6.196.439.9810.45
Appendix
Table 9: Effect of the base scoring function on UI-TARS-1.5-7B / full AITW ( 4,6944{,}694 step instances) at 20%20\% budget. CSS and TSG provide consistent gains on top of both bases tested; the scheme’s effectiveness does not depend on either specific AbaseA_{base} choice.
CategoryNBase onlyST-Lite (CSS+TSG)Δ\Delta
general84215.020.5+5.5+5.5
googleapps50513.418.4+5.0+5.0
install125317.719.3+1.6+1.6
single42215.821.7+5.9+5.9
webshopping167214.220.5+6.3+6.3
All469415.320.1+4.8\mathbf{+4.8}
Appendix
Table 10: Per-category step accuracy (%) on AITW at 20% budget, UI-TARS-1.5-7B. Δ\Delta = absolute improvement over SnapKV base. The gain is largest on webshopping ( +6.3+6.3 ) and smallest on install ( +1.6+1.6 ).
LengthNBaseFullΔ\Delta
1-3 steps27614.016.5+2.5+2.5
4-6 steps63816.220.3+4.1+4.1
7-9 steps114816.120.5+4.4+4.4
10-13 steps128115.220.2+5.0+5.0
14 ++ steps135114.520.4+5.9+5.9
Appendix
Table 11: Step accuracy (%) on AITW at 20% budget, by parent-episode length. Δ\Delta grows monotonically from +2.5%+2.5\% on short episodes to +5.9%+5.9\% on the longest.
α\alphaTSG = offTSG = on
0 (base/TSG only)15.318.9
0.516.219.9
1.017.420.7
2.016.719.6
Appendix
Table 12: Panel (a): α×\alpha\times TSG sensitivity on the full AITW corpus, UI-TARS-1.5-7B, 20%20\% KV budget, full-attention base scoring. TSG-on dominates TSG-off at every α\alpha by +2.9+2.9 to +3.7%+3.7\% ; CSS contributes +0.9+0.9 to +2.1%+2.1\% across both TSG settings. The joint optimum is α=1\alpha=1 , TSG=on at 20.7%20.7\% . Per-step aggregation under the full-attention base scorer; the per-step aggregation reported in Table 38 is 20.1%20.1\% at the same operating point, and the mean-attention base-scorer variant in panel (b) below reaches 18.19%18.19\% . Bold marks the panel maximum.
Basebase only+CSS+TSG+CSS+TSG
uniform6.196.43 ( +0.24+0.24 )9.98 ( +3.79+3.79 )10.45 ( +4.26+4.26 )
mean_attn16.4718.12 ( +1.65+1.65 )17.89 ( +1.42+1.42 )18.19 ( +1.72+1.72 )
Appendix
Table 13: Panel (b): module decomposition on the full 4,6944{,}694 -step AITW corpus at 20%20\% KV budget under two base scorers. Each row reports a 2×22{\times}2 ablation; parenthesised values are absolute % gains over the row’s base-only cell. Bold marks the row maximum.
Budget alloc.Step Acc. (%)CHR (%)
Flat (default)20.166.1
Hier. (PyramidKV)17.356.8
Hier. (VL-Cache)16.954.3
Appendix
Table 14: Budget allocation ablation within ST-Lite on AITW (UI-TARS-1.5-7B, β=20%\beta{=}20\% , 4,694 steps). Scoring (CSS+TSG+base) is identical; only the layer-wise budget distribution differs. Flat allocation gains +2.8+2.8 – 3.2%3.2\% accuracy and +9.3+9.3 – 11.8%11.8\% CHR over hierarchical variants, confirming P3 .
Budget (%)Step acc. (%)Regime
313.2extreme compression
516.0steep growth
1018.4approaching peak
1520.0near-peak
2020.1less-is-more peak
4019.0post-peak plateau
Appendix
Table 15: Representative anchors from a fine-grained KV-cache budget sweep for ST-Lite on the full AITW evaluation set (UI-TARS-1.5-7B, δ=8\delta=8 , 4,694 step instances). The 15-budget sweep in [3,100]%[3,100]\% confirms a single smooth curve; the six rows below isolate the qualitatively distinct regimes.
Budget (%)click acc. (%)type acc. (%)
517.210.5
1019.813.1
2021.415.2
4020.316.8
6019.716.1
Appendix
Table 16: Per-action-type accuracy (%) of ST-Lite against budget on the full AITW evaluation set (UI-TARS-1.5-7B, δ=8\delta=8 , 4,694 step instances). Only the two dominant action types are shown ( click ≈70%\approx 70\% , type ≈12%\approx 12\% ).
BenchmarkMethodβ=1%\boldsymbol{\beta{=}1\%}β=3%\boldsymbol{\beta{=}3\%}β=5%\boldsymbol{\beta{=}5\%}β=10%\boldsymbol{\beta{=}10\%}Full
ScreenSpotV2SnapKV16.162.977.585.288.9
MixKV17.063.178.385.888.9
ST-Lite17.162.979.686.088.9
AITWSnapKV8.111.513.214.218.2
MixKV8.311.813.514.618.2
ST-Lite8.613.216.018.418.2
Appendix
Table 17: Comparison with MixKV under extreme compression ( β≤10%\beta\leq 10\% ) on UI-TARS-1.5-7B. ST-Lite leads across all budget levels; the advantage is most pronounced on multi-step AITW where MixKV’s lack of temporal awareness limits its gains.
Figure 9: Token retention on a 5-step flight-search trajectory (“Mexico City to Zurich”) at β=20%\beta{=}20\% . Highlighted regions = retained tokens; dark = evicted. Methods: SnapKV , PyramidKV , VL-Cache , ST-Lite , across historical and current frames.
Window size δ\delta20% budget40% budget
415.3217.81
8 (default)20.1019.03
1618.7419.42
3216.2118.87
6417.4318.14
Appendix
Table 18: Step accuracy (%) on the full AITW evaluation set (4,694 step instances, UI-TARS-1.5-7B), sweeping the observation-window size δ\delta at two budgets. The default δ=8\delta=8 inherited from SnapKV is the best setting at the tight 20%20\% budget and remains competitive at the more relaxed 40%40\% budget.
Similarity MetricAccuracy (%)
Cosine (default)19.1
Dot Product18.3
Negative L2 Distance18.2
Appendix
Table 19: TSG similarity-metric ablation under a fair comparison: identical base scoring (Mean Attention), identical 20% budget, observation window δ=16\delta{=}16 . Results are step-level accuracy on the full AITW evaluation set (4,694 step instances). The slightly higher absolute accuracy relative to Table 13 panel (b) ( 19.1%19.1\% compared with 18.19%18.19\% ) reflects the larger observation window ( δ=16\delta{=}16 rather than δ=8\delta{=}8 ); within this table the three metrics agree within ≤\leq 1%, indicating that TSG’s effectiveness comes from the adaptive thresholding mechanism rather than the choice of similarity function. Cosine remains our default for the reasons discussed below.
Nbhd. k×kk{\times}kStep Acc. (%)CHR (%)
3×33\times 3 (default)20.166.1
5×55\times 518.659.7
7×77\times 719.061.2
9×99\times 917.857.3
Appendix
Table 20: CSS neighborhood size ablation on AITW (UI-TARS-1.5-7B, β=20%\beta{=}20\% , 4,694 steps). The 3×33{\times}3 kernel achieves the best accuracy and CHR.
ConfigurationStep Acc. (%)CHR (%)
ST-Lite (adaptive TSG + Laplacian CSS)20.166.1
TSG controls (CSS fixed at Laplacian, α=1\alpha{=}1 )
Random temporal gate (same drop rate)17.054.3
Fixed threshold ( τ=0.85\tau{=}0.85 )18.458.7
TSG off ( α=1\alpha{=}1 , CSS only)17.455.8
CSS controls (adaptive TSG fixed)
Appendix
Table 21: Component necessity controls on AITW (UI-TARS-1.5-7B, β=20%\beta{=}20\% , 4,694 steps). Random temporal gate : for each step, randomly mask the same fraction of historical frames that adaptive TSG would have masked (matched drop rate). Fixed threshold : gate all frames with cosine similarity >0.85>0.85 regardless of distribution. Random spatial scores : replace Laplacian uniformity with i.i.d. U[0,1]\mathcal{U}[0,1] per token. Gaussian-blur spatial : replace Laplacian with a 3×33{\times}3 Gaussian-smoothed hidden-state norm (destroys boundary signal while retaining spatial smoothness).
CategoryMean Φspace\Phi_{space}Pctl.NN /img
Boundary0.34178.397
Interior0.07228.643
Background0.10847.11,156
Appendix
Table 22: CSS saliency by token category on ScreenSpot Pro ( 100100 images, 1,5811{,}581 elements). Pctl. = average percentile rank of Φspace\Phi_{space} among all 1,2961{,}296 visual tokens per image. Boundary tokens receive 4.7×4.7{\times} higher CSS scores than interior tokens and rank at the 7878 th percentile on average.
Top- BB %Prec.RecallF1×\times rand.
5%0.520.350.426.9×6.9\times
10%0.420.560.485.6×5.6\times
20%0.300.800.444.0×4.0\times
30%0.220.910.352.9×2.9\times
Appendix
Table 23: CSS boundary precision/recall on ScreenSpot Pro. Random precision =7.5%=7.5\% . At the deployment budget ( B=20%B{=}20\% ), CSS captures 80%80\% of boundary tokens at 4×4{\times} random precision.
ScoringBnd. pctl.Bnd. Recall@20%
Random50.020.0%
AbaseA_{base} only54.226.3%
CSS ( Φspace\Phi_{space} )78.380.4%
Abase+αΦspaceA_{base}{+}\alpha\Phi_{space}71.668.7%
Appendix
Table 24: Boundary-token ranking under different scoring functions. AbaseA_{base} is barely above chance at placing boundary tokens high; CSS provides the dominant boundary signal.
QuadrantCSSAbaseA_{base}CombinedRetention
 (mean)(mean)(mean)@20%
High-CSS + Task-relevant0.3580.4120.77088.3%
High-CSS + Task-irrelevant0.3410.0890.43034.7%
Low-CSS + Task-relevant0.0710.3870.45871.2%
Low-CSS + Task-irrelevant0.0630.0520.1154.8%
Appendix
Table 25: Task-relevance filtering by combined scoring on ScreenSpot Pro ( β=20%\beta{=}20\% , UI-TARS-1.5-7B). Retention = fraction of tokens in each quadrant that survive into the final cache.
CategoryMean Φspace\Phi_{space}Pctl.NN /img
Boundary0.32776.184
Interior0.08131.238
Background0.11248.31,174
Appendix
Table 26: CSS saliency by token category on AITW current frames (200 multi-frame steps, UI-TARS-1.5-7B). Boundary tokens of the action target receive 4.0×4.0{\times} higher CSS scores than interior tokens (cf. 4.7×4.7{\times} on ScreenSpot Pro, Table 22 ), confirming that the spatial-boundary signal persists in the multi-frame regime.
Top- BB %Prec.RecallF1×\times rand.
5%0.400.310.356.2×6.2\times
10%0.330.510.405.1×5.1\times
20%0.240.740.363.7×3.7\times
30%0.190.870.312.9×2.9\times
Appendix
Table 27: CSS boundary precision/recall on AITW current frames. Random precision = 84/1296=6.5%84/1296=6.5\% . At the deployment budget ( B=20%B{=}20\% ), CSS captures 74%74\% of action-target boundary tokens at 3.7×3.7{\times} random precision (cf. 80%80\% recall, 4.0×4.0{\times} on ScreenSpot Pro, Table 23 ). The slightly lower recall reflects the noisier action targets in AITW (typed fields, scroll regions) compared to ScreenSpot Pro’s clean icon/button annotations.
ParadigmTrainParam.Operates onGoal
Std. KV cacheNoNoFull KVStd. inference
PEFT / LoRAYesYes (LR)Model paramsAdaptation
Test-time adapt.OnlineYesParams / statsDist. shift
ST-LiteNoNoKV cacheMemory / lat.
Appendix
Table 28: Positioning of ST-Lite relative to neighbouring adaptation paradigms.
 ScreenSpotv2AndroidControl (High)AndroidControl (Low)
Methodβ=10%\beta{=}10\%β=20%\beta{=}20\%β=40%\beta{=}40\%β=10%\beta{=}10\%β=20%\beta{=}20\%β=40%\beta{=}40\%β=10%\beta{=}10\%β=20%\beta{=}20\%β=40%\beta{=}40\%
Full Cache73.373.373.348.448.448.462.462.462.4
SnapKV61.568.171.838.242.746.549.855.960.3
PyramidKV55.863.069.734.539.444.145.151.357.8
VL-Cache54.261.869.033.138.043.343.750.056.9
GUI-KV63.069.572.239.844.647.452.458.361.5
Appendix
Table 29: Cross-architecture validation on UGround-7B (LLaVA-NeXT backbone). Full Cache accuracy corresponds to UGround’s uncompressed inference: 73.3%73.3\% on ScreenSpot (averaged over platforms), 48.4%48.4\% on AndroidControl high-level, and 62.4%62.4\% on AndroidControl low-level (GPT-4o planner, 500 test steps). ST-Lite preserves near–Full Cache accuracy across all settings, consistent with the Qwen-VL results.
MethodPrefillDecodeTotalTputE2ETput
 (ms)(ms)(ms)(tok/s)speedupspeedup
Full Cache (100%)241784941091125.11.00 ×\times1.00 ×\times
SnapKV (20%)24325153758533.51.44 ×\times1.34 ×\times
PyramidKV (20%)24557446990135.31.10 ×\times1.40 ×\times
VL-Cache (20%)24666972943835.01.16 ×\times1.39 ×\times
GUI-KV (20%)31024313741536.81.47 ×\times1.47 ×\times
Appendix
Table 30: Per-method inference latency on the complete AgentNetBench evaluation set at image_slots =10, single 4848 GB GPU. All non-Full-Cache methods use a 20% KV budget. Speedup is relative to Full Cache; values >1>1 are faster. The numbers here are consistent with the main-text microbenchmark in Table 2 at 10 screenshots.
Traj. lengthN stepsFC dec (ms)ST dec (ms)Decode speedupE2E speedup
16321008836822.74 ×\times2.04 ×\times
1818735536432.02 ×\times1.61 ×\times
1919826835412.34 ×\times1.79 ×\times
2323735335712.06 ×\times1.63 ×\times
All92849436182.35 ×\times1.80 ×\times
Appendix
Table 31: ST-Lite decoding and end-to-end speedup over Full Cache stratified by parent-trajectory length (UI-TARS-1.5-7B, AgentNetBench, 20% budget, single 4848 GB GPU). The decoding speedup remains ≥2×\geq 2\times across every length bin and peaks at 2.74×2.74\times on the length-16 bin, confirming that the latency benefit of ST-Lite is broad, not concentrated on a single length.
MethodPre.Dec.E2ETput
Full Cache1.00 ×\times1.00 ×\times1.00 ×\times7.01
ST-Lite (20%)0.98 ×\times1.94 ×\times1.48 ×\times11.29
Appendix
Table 32: Latency and throughput of ST-Lite on OpenCUA-32B (complete AgentNetBench evaluation set, 20%20\% KV cache budget, 1010 image slots). Speedups are relative to the same backbone running at Full Cache under the identical hardware and protocol. The relative speedup is consistent with the 7B numbers in Table 30 (decode 2.35×→1.94×2.35\times\to 1.94\times , end-to-end 1.80×→1.48×1.80\times\to 1.48\times ), confirming that ST-Lite scales gracefully to substantially larger backbones. Tput is in tok/s.
Figure 10: Less-is-More on full AITW (UI-TARS-1.5-7B). ST-Lite is the only plotted compression method with a sustained above-Full-Cache region across an interior budget range, peaking at β=20%\beta{=}20\% with a +2.96%+2.96\% gap; SnapKV and PyramidKV only marginally brush the Full Cache line at isolated higher budgets (PyramidKV at β=80%\beta{=}80\% by +0.04%+0.04\% , within per-episode reporting noise; SnapKV at β=40%\beta{=}40\% by +0.45%+0.45\% , above noise but at a single budget and far below ST-Lite’s peak). The Full Cache reference ( β=100%\beta{=}100\% , 17.14%17.14\% ) is shown as the horizontal dashed line rather than a rightmost point.
CategoryFCSnapST-LiteΔST-FC\Delta_{\text{ST-FC}}
 @100%@20%@20% 
general15.415.721.5+6.1+6.1
single18.515.621.6+3.1+3.1
webshopping16.016.119.6+3.5+3.5
install19.518.221.2+1.7+1.7
googleapps16.010.916.6+0.6+0.6
Appendix
Table 33: AITW action accuracy (%) by task category, UI-TARS-1.5-7B. ST-Lite at β=20%\beta{=}20\% beats the Full Cache reference on every category, and beats SnapKV at the same budget by +3.0+3.0 to +6.0%+6.0\% .
β\betaNNBothST-onlySn-onlyNeitherNET
3%3\%46943123082283846+80+80
5%5\%46943474042733670+131+131
10%10\%46943615033063524+197+197
15%15\%46944225172593496+258+258
20%20\%46704574822583473+224+224
40%40\%45634803872323464+155+155
Appendix
Table 34: Paired ST-Lite-vs-SnapKV step-level decomposition on AITW (UI-TARS-1.5-7B). NET (ST-Lite unique wins −- SnapKV unique wins) is positive at every budget and peaks at +258+258 at β=15%\beta{=}15\% ; the gap collapses to noise only at β=80%\beta{=}80\% , the signature of decision-relevant retention. Per-method accuracies reproduce main-body Table 38 .
Figure 11: (a) Paired step-level decomposition on AITW. Bars stack three outcomes (both/only-ST-Lite/only-SnapKV) as % of N≈4,500N{\approx}4{,}500 common steps; NET label = ST-Lite unique wins minus SnapKV unique wins. (b) The NET gap peaks in β∈[10%,20%]\beta\in[10\%,20\%] and collapses to noise at β=80%\beta{=}80\% .
TaskNNBothOnly STOnly SnNeitherNET
general842849748613+49+49
webshopping16581531711141220+57+57
install124314512181896+40+40
googleapps505315324397+29+29
single422444722309+25+25
Appendix
Table 35: Paired ST-Lite-vs-SnapKV decomposition at β=20%\beta{=}20\% , stratified by AITW task category. NET is positive on every category.
Figure 12: Three long-horizon AITW trajectories on which ST-Lite is correct and SnapKV is not, from the three categories with the largest paired NET gap in Table 35 . Green borders: ST-Lite uniquely correct. Italic captions: trajectory-level outcome.
Figure 13: Webshopping trajectory at β=20%\beta{=}20\% (UI-TARS-1.5-7B). Green border: ST-Lite uniquely correct. Grey border: tie. The three ST-only wins (steps 3/4/7) are the cart-clearing and search-result steps that depend on context outside SnapKV’s δ=8\delta{=}8 observation window.
β\betaST-LiteSnapKVPyr.KVVL-Cache
357.757.643.432.0
574.071.952.453.1
1079.979.169.965.6
1580.480.977.277.8
2083.382.878.279.7
4083.483.082.282.8
Appendix
Table 36: Icon-grounding accuracy (%) on ScreenSpot-V2 across the budget sweep. Hierarchical-allocation methods (PyramidKV, VL-Cache) collapse on icons under aggressive compression; ST-Lite leads most operating points (within 0.5%0.5\% of the best at β=15%\beta{=}15\% and β=80%\beta{=}80\% ).
Figure 14: Element-type decomposition of ScreenSpot-V2 grounding accuracy under KV cache compression. Left: text-target accuracy is similar across methods once β≥10%\beta\geq 10\% . Right: icon-target accuracy diverges sharply at low budgets — hierarchical methods drop ≥20%\geq 20\% behind ST-Lite at β=5%\beta{=}5\% on icons, while the text gap is far smaller. Empirical signature of structure misalignment.
Figure 15: Cross-platform robustness on ScreenSpot-V2 under aggressive KV-cache compression. Each panel reports combined grounding accuracy by platform at the labelled budget. ST-Lite dominates every platform-budget cell.
MethodMobileDesktopWeb
ST-Lite88.070.380.1
SnapKV84.868.779.2
PyramidKV53.953.459.2
VL-Cache64.726.868.1
Δ\Delta over PyramidKV+34.1+34.1+16.9+16.9+20.9+20.9
Δ\Delta over VL-Cache+23.3+23.3+43.4+43.4+12.0+12.0
Appendix
Table 37: ScreenSpot-V2 grounding accuracy (%) at β=5%\beta{=}5\% , by platform. ST-Lite leads all three platforms; VL-Cache regresses on desktop.
More in this topic

KV-Cache Compression

Key-value cache compression reduces the memory and bandwidth needed to retain attention states during long-context inference.

See all 267 papers

Explore similar work

CardsList

  1. STaR-KV: Spatio-Temporal Adaptive Re-weighting for KV Cache Compression in GUI Vision-Language Models

    Jun 1, 2026Yuhang Han, Wenzheng Yang, Yujie Chen +4Vision-Language ModelsKV Caching

    1EPIC Lab, SJTU · 2HKUST (GZ) · 3The University of Sydney +3

  2. GUI-KV: Efficient GUI Agents via KV Cache with Spatio-Temporal Awareness

    Oct 1, 2025Kung-Hsiang Huang, Haoyi Qiu, Yutong Dai +2KV-Cache CompressionEfficient Inference

    Salesforce AI Research · University of California, Los Angeles

  3. Spatio-Temporal Token Pruning for Efficient High-Resolution GUI Agents

    Feb 26, 2026Zhou Xu, Bowen Zhou, Qi Wang +2GUI AgentsVisual Token Pruning

    Tsinghua University, Shenzhen, China · Xidian University, Xi’an, China · The Chinese University of Hong Kong, Hong Kong, China