GUIOdyssey: A Comprehensive Dataset for Cross-App GUI Navigation on Mobile Devices (OdysseyAgent)

P001 原表

来源：作者原文。2026-09-07抓取，原始行列，不是统一复跑成绩。

Table 1 : GUI navigation dataset comparison. GUIOdyssey is a comprehensive cross-app GUI navigation dataset with over 8 8 k+ episodes, featuring an average of 15.3 15.3 steps, which is the longest among mobile GUI datasets, and includes a diverse range of devices such as tablets.

Dataset	# Episodes	# Unique $\bm{I_{user}}$	# Avg. Steps	Cross-app?	Platform	# Domains	Instruction Level	Semantic Annotation
Mind2Web [17]	2,350	2,350	7.3	✗	Web	137 sites	high	✗
WebLINX [32]	2,337	2,337	43.0	✗	Web	155 sites	high	✗
PixelHelp [27]	187	187	4.2	✗	Phone	4 apps	high & low	✗
MoTIF [8]	4,707	276	4.5	✗	Phone	125 apps	high & low	✗
UGIF [47]	523	480	5.3	✗	Phone	12 apps	high & low	✓
Meta-GUI	4,684	1,125	5.3	✗	Phone	11 apps	high	✗
AITW [66]	715,142	30,378	6.5	✗	Phone	159 apps, 198+ sites	high	✗
AITZ [66]	2,504	2,504	7.5	✗	Phone	70+ apps	high	✓
AndroidControl [26]	15,283	14,548	5.5	✗	Phone	833 apps	high & low	✗
AMEX [9]	2,946	2,946	12.8	✗	Phone	110 apps	high	✓
GUIOdyssey	8,334	8,334	15.3	✓	Phone & Tablet	212 apps, 1,357 app combos	high & low	✓

Table 2 : Results of different LVLMs on Test-Random split. The evaluation metric is the action matching score (AMS). ‘HL’ and ‘LL’ indicate that the task instruction is high-level and low-level, respectively. ∗ indicates that agent’s training also includes semantic annotations.

zero-shot
Model	Tool		Information		Shopping		Media		Social		Multi-Apps		Overall
Model	HL	LL	HL	LL	HL	LL	HL	LL	HL	LL	HL	LL	HL	LL
GPT-4V	14.93	40.86	14.69	38.56	12.17	36.04	10.80	48.40	16.79	43.21	11.54	40.61	13.49	41.28
GPT-4o	14.15	38.11	13.86	42.40	11.69	40.57	12.00	52.40	18.14	44.44	9.28	38.35	13.19	42.71
Claude3.5-sonnet	22.99	40.28	14.69	34.56	12.17	31.03	14.00	37.60	16.79	30.62	14.14	31.00	15.80	34.18
InternVL2-Pro	19.45	49.51	15.86	40.23	17.18	41.05	13.20	51.60	14.81	40.00	15.72	41.52	16.04	43.98
CogAgent	18.81	33.52	12.35	29.79	13.02	26.89	12.63	25.80	14.72	33.54	12.15	33.09	13.95	30.44
SphAgent	22.24	36.81	17.43	29.95	13.60	25.08	15.54	33.03	14.20	28.83	12.89	26.32	15.98	30.00
zero-shot with OmniParser
GPT-4V	24.37	56.03	22.44	51.10	17.16	46.75	18.65	62.18	32.28	59.18	24.21	54.58	23.18	54.97
GPT-4o	26.63	55.28	23.45	53.91	18.34	47.63	19.17	63.21	31.01	61.08	23.39	55.95	23.67	56.18
Claude3.5-sonnet	39.20	64.07	28.46	61.92	27.22	56.51	28.50	66.84	40.82	68.99	33.11	65.12	32.88	63.91
InternVL2-Pro	16.58	58.04	16.03	51.30	8.88	49.70	16.58	60.62	16.14	53.80	13.95	52.39	14.69	54.31
fine-tuned
Qwen-VL	85.55	90.79	68.04	83.36	62.28	80.67	77.56	88.15	80.29	88.36	74.27	86.56	74.67	86.32
${\text{Qwen-VL}}^{\ast}$	86.35	90.99	72.01	85.77	67.31	82.86	80.33	89.48	82.39	88.01	77.52	89.40	77.65	87.78
OdysseyAgent	86.01	91.21	69.83	83.37	65.19	82.63	77.10	88.55	81.47	87.66	75.13	87.84	75.79	86.88
${\textbf{OdysseyAgent}}^{\ast}$	86.82	91.25	71.79	86.58	68.58	83.74	80.93	89.66	82.88	88.27	78.47	89.39	78.24	88.15

Table 3 : OdyssseyAgent’s performance on out-of-domain tasks. ‘Semantic?’ indicates whether the model’s training includes semantic annotations. ‘HL’ and ‘LL’ indicate that the task instruction is high-level and low-level, respectively.

Semantic?	Task	Test-Task		Test-Device		Test-App		Overall
Semantic?	Level	AMS	SR	AMS	SR	AMS	SR	AMS	SR
✗	HL	54.36	0.09	61.20	1.88	63.03	7.70	59.53	3.22
✗	LL	78.97	2.20	79.66	8.47	84.24	20.70	80.96	10.46
✓	HL	56.19	0.26	66.63	5.07	65.89	8.81	62.90	4.71
✓	LL	80.19	2.29	79.93	11.66	83.47	20.02	81.20	11.32

Table 4 : The impact of different historical components in GUIOdyssey across four splits. High-level instructions are used for both training and evaluation, and performance is measured by AMS and SR.

	Historical Information			Test-Random		Test-Task		Test-Device		Test-App		Overall
	action	screenshot	context	AMS	SR	AMS	SR	AMS	SR	AMS	SR	AMS	SR
(1)	✗	✗	✗	66.13	1.65	47.62	0.00	54.15	0.72	54.49	3.59	55.60	1.49
(2)	✓	✗	✗	74.67	9.70	55.00	0.00	62.03	2.03	62.06	8.98	63.44	5.18
(3)	✗	✓	✗	71.22	6.69	51.69	0.09	59.12	2.24	59.16	7.78	60.30	4.20
(4)	✗	✗	✓	75.25	9.50	57.66	0.62	62.35	2.24	63.82	7.87	64.77	5.06
(5)	✓	✓	✗	75.79	9.38	54.36	0.09	61.20	1.88	63.03	7.70	63.60	4.76
(6)	✓	✓	✓	77.06	11.61	58.83	0.18	65.85	5.00	65.63	8.47	66.84	6.32

Table 5 : The argument and functionality of different actions in GUIOdyssey. ‘pos1’ and ‘pos2’ denote the position ( x , y ) (x,y) .

Action	Argument	Functionality
CLICK	[pos1]	click the on-screen position
LONG PRESS	[pos1]	press the screen for a long time to copy texts or download images
SCROLL	[pos1, pos2]	scroll the screen from position 1 to position 2
TYPE	text	type text with keyboard
COMPLETE	-	the sign that the instruction has been completed
IMPOSSIBLE	-	the sign that the instruction cannot be completed
HOME	-	go to the home screen
BACK	-	go to the previous screen
RECENT	-	go to the previous App

Table 6 : The impact of different semantic annotations on OdysseyAgent across four different splits. We use high-level instructions for both training and evaluation. Performance is assessed using AMS and SR as metrics. SD, CI, and DR denote screen description, contextual information, and decision rationale, respectively.

	Semantic Annotation			Test-Random		Test-Task		Test-Device		Test-App		Overall
	SD	CI	DR	AMS	SR	AMS	SR	AMS	SR	AMS	SR	AMS	SR
(1)	✗	✗	✗	75.79	9.38	54.36	0.09	61.20	1.88	63.03	7.70	63.60	4.76
(2)	✓	✗	✗	75.18	8.94	54.06	0.00	64.41	2.03	64.91	8.47	64.64	4.86
(3)	✗	✓	✗	75.42	10.04	55.71	0.00	62.52	3.19	64.24	5.30	64.47	4.63
(4)	✗	✗	✓	77.71	11.44	55.60	0.26	65.88	4.63	65.74	7.96	66.23	6.07
(5)	✗	✓	✓	77.23	11.16	56.93	0.18	63.87	2.24	66.32	7.87	66.09	5.36
(6)	✓	✗	✓	77.24	10.88	57.15	0.00	63.55	2.17	67.04	9.67	66.24	5.68
(7)	✓	✓	✗	76.58	10.14	57.13	0.26	64.48	3.91	66.27	7.96	66.11	5.57
(8)	✓	✓	✓	78.24	11.62	56.19	0.26	66.63	5.07	65.89	8.81	66.74	6.44

Table 7 : The average AMS for HL and LL instructions across 4 splits, along with the number of historical screenshot tokens, inference metrics (Time to First Token (TTFT) and Tokens per Second (TPS)), and training GPU hours.

strategy	HL	LL	Token Count	TTFT $\downarrow$	TPS $\uparrow$	GPU Hours
history resampler	63.60	82.44	256	0.71	20.27	32
multi-image	65.04	82.34	1024	0.98	17.05	48

Table 8 : The results for OdysseyAgent trained and tested on Train-Random/Test-Random with both high-level and low-level instructions are presented, with AMS as the evaluation metric. HL and LL denote high-level and low-level instructions, respectively.

Testing Instructions	Training Instructions
Testing Instructions	HL	LL	HL + LL
HL	75.79	29.39	78.96
LL	71.26	86.88	88.84

Table 9 : Performance Evaluation of OdysseyAgent Across Different Devices. Each Device serves as a test set while the remaining five devices are used as training sets.

Evaluation Device	Resolution	AMS	SR
Pixel 7 Pro	$1,440\times 3,120$	75.91	7.44
Pixel 8 Pro	$1,344\times 2,992$	74.67	6.05
Small Phone	$720\times 1,280$	71.68	3.77
Medium Phone	$1,080\times 2,400$	73.05	5.45
Pixel Fold	$2,208\times 1,840$	67.67	4.48
Pixel Tablet	$2,560\times 1,600$	61.20	1.88

Table 10 : Effectiveness of Different Training Data on the AndroidControl. The evaluation metrics are the action matching score (AMS).

Training Data	IDD	category_unseen	app_unseen	task_unseen	Overall
AC50k	60.43	54.46	50.00	72.10	59.25
AC50k + AITW50k	60.69	55.26	45.19	68.84	57.50
AC50k + Ody50k	61.48	54.61	50.96	72.46	59.88