Benchmarking and Improving GUI Agents in High-Dynamic Environments (DynamicGUIBench / DynamicUI) 原表

Y26-097 原文表格

抽取日期：2026-09-07。来自原文。保留原表行列，未统一实验配置；数值解释见中文分析。

Table 1. Comparison of existing GUI benchmarks and our benchmark in terms of application diversity, task scale, evaluation setting (Mode), anomaly presence (An.), and dynamic tasks (Dyn.). ✓ and ✗ denote the presence and absence of anomaly or dynamic tasks, respectively.

Benchmark	Venue	Apps/Web	Tasks	Mode	An.	Dyn.
Mind2Web (Deng et al., 2023)	NeurIPS’23	137	2,350	Offline	✗	✗
OSWorld (Xie et al., 2024)	NeurIPS’24	10	369	Online	✗	✗
GUI Odyssey (Lu et al., 2025)	ICCV’25	201	7,735	Offline	✗	✗
WorldGUI (Hengyuan Zhao et al., 2025)	arXiv’25	10	315	Offline	✗	✗
Android World (Rawles et al., )	ICLR’25	20	116	Offline	✗	✗
OSWorld-G (Xie et al., 2025)	NeurIPS’25	10	564	Online	✗	✗
GUI-Robust (Yang et al., 2025b)	NeurIPS’25	392	5,318	Offline	✓	✗
D-GARA (Chen et al., 2025a)	AAAI’26	7	152	Online	✓	✗
Ours	2026	10	149	Online	✓	✓

Table 2. Key statistics in DynamicGUIBench.

Statistic	Number	Statistic	Number
Chrome	38 (25.5%)	Multi apps	33 (22.4%)
Gimp	10 (6.8%)	Os	7 (4.7%)
Libreoffice ${}_{\text{calc}}$	14 (9.5%)	Thunderbrid	9 (6.1%)
Libreoffice ${}_{\text{impress}}$	7 (4.7%)	VLC	5 (3.4%)
Libreoffice ${}_{\text{writer}}$	14 (9.5%)	VS Code	12 (8.1%)
Feasible	146 (98%)	Infeasible	3 (2.0%)

Table 3. Comparison with state-of-the-art methods on the DynamicGUIBench. Abbreviations: Th. (Thunderbird), Multi. (Multi-Apps), Vs. (Visual Studio Code), Imp. (Impress), Wri. (Writer). The best scores are in bold.

Agent Model	Step	Acc (%)
Agent Model	Step	Chrome	Gimp	Calc	Imp.	Wri.	Multi.	Os	Th.	Vlc	Vs.	Avg.
o3 (OpenAI, 2025)	15	7.9	10.0	7.1	0.0	7.1	6.1	0.0	0.0	20.0	8.3	6.7
Qwen3-vl-4B (Bai et al., 2025)	15	18.4	10.0	7.1	0.0	7.1	9.1	14.3	0.0	40.0	0.0	10.7
Qwen3-vl-8B (Bai et al., 2025)	15	21.1	20.0	14.3	0.0	14.3	9.1	14.3	0.0	40.0	16.7	14.8
o3 (OpenAI, 2025)	50	10.5	0.0	14.3	0.0	7.1	6.1	42.9	0.0	20.0	0.0	8.7
UITARS-1.5-7B (Qin et al., 2025b)	50	13.2	10.0	7.1	0.0	7.1	9.1	0.0	0.0	20.0	0.0	8.1
Qwen3-vl-4B (Bai et al., 2025)	50	13.2	10.0	7.1	0.0	7.1	9.1	28.6	0.0	60.0	0.0	10.7
Qwen3-vl-8B (Bai et al., 2025)	50	26.3	22.0	14.3	0.0	7.1	7.5	14.3	0.0	60.0	8.3	15.1
doubao-1-5-0717(Guo et al., 2025)	50	5.3	0.0	7.1	0.0	0.0	0.0	0.0	0.0	0.0	0.0	2.8
Seed1.8-VL (Seed, )	50	10.5	10.0	0.0	14.3	7.1	3.0	0.0	0.0	0.0	16.7	6.7
EvoCUA-8B (Xue et al., 2026)	50	18.4	33.0	0.0	0.0	7.1	4.5	28.6	0.0	20.0	16.7	11.7
Ours w/ Qwen3-vl-8B	15	29.0	10.0	14.3	0.0	7.1	12.2	14.3	11.1	0.0	16.7	15.5
Ours w/ Qwen3-vl-8B	50	36.8	30.0	14.3	14.3	7.1	9.1	28.6	44.4	40.0	8.3	22.1

Table 4. Ablation on the DynamicGUIBench.

Dynamic Perceiver	Reflection Module	Refinement Strategy	Acc(%)
-	-	-	15.1
✓	-		17.4
✓	-	✓	17.4
✓	✓	-	20.8
✓	✓	✓	22.1

Table 5. Comparison with state-of-the-art methods on the OSWorld ( Xie et al., 2024 ) benchmark. We use ‘*’ to denote the results evaluated by us (which will be updated if improved evaluation scripts become available).

Agent Model	Step	Acc (%)
Agent Model	Step	Chrome	Gimp	Calc	Imp.	Wri.	Multi.	Os	Th.	Vlc	Vs.	Avg.
o3 (OpenAI, 2025)	50	21.7	38.5	8.5	4.3	21.7	11.8	37.5	20.0	17.6	21.7	17.2
Opencua-a3b (Wang et al., 2025a)	50	21.6	53.9	2.1	23.3	34.8	6.5	37.5	6.7	11.8	43.5	19.9
UITARS-72b-dpo (Qin et al., 2025b)	50	33.2	61.5	12.8	25.5	43.5	6.7	33.3	33.3	23.5	47.8	25.8
UITARS-1.5-7B (Qin et al., 2025b)	50	32.8	53.9	8.5	39.3	41.3	8.6	25.6	46.7	24.4	52.2	27.2
Qwen3-vl-4B* (Bai et al., 2025)	50	30.4	21.7	12.8	21.8	39.1	5.2	33.3	46.7	23.0	13.0	19.8
Opencua-7B (Wang et al., 2025a)	50	38.6	43.6	13.2	32.6	33.3	12.1	43.5	42.2	28.3	47.1	28.2
Qwen3-vl-8B* (Bai et al., 2025)	50	23.9	34.8	17.0	31.8	56.5	9.2	33.3	66.7	40.4	21.7	25.8
Ours w/ UITARS-1.5-7B	50	32.8	57.7	8.5	39.3	41.3	8.6	29.8	50.0	27.5	54.4	28.2
Ours w/ Qwen3-vl-8B	50	34.8	69.6	17.0	19.6	26.1	13.1	62.5	40.0	17.7	43.5	28.4

Table 6. Comparison of uniform frame sampling baselines and our Dynamic Perceiver (DP) across four POMDP categories on DynamicGUIBench.

Method	Frames	Interruptive UI	Ephemeral Ref.	DynList	ContentTrig	Avg.
Qwen3-vl-8B (Bai et al., 2025)	-	25.0	12.3	13.6	19.2	15.1
Uniform	1	25.0	15.3	31.8	3.9	16.8
Uniform	3	25.0	17.0	13.6	11.5	16.4
Ours	DP	43.8	20.0	22.7	15.4	22.1

Table 7. Comparison of different general-purpose models for the reflection module on DynamicGUIBench.

Method	Interruptive UI	Ephemeral Ref.	DynList	ContentTrig	Avg.
Qwen3-vl-8B (Bai et al., 2025)	25.0	9.3	9.1	11.5	11.3
GPT-5.4-mini (Singh et al., 2025)	43.8	20.0	22.7	15.4	22.1