We introduce WebChain, the largest open-source dataset of human-annotated trajectories on real-world websites, designed to accelerate reproducible research in web agents. It contains 31,725 trajectories and 318k steps, featuring a core Triple Alignment of visual, structural, and action data to provide rich, multi-modal supervision. The data is collected via a scalable pipeline that ensures coverage of complex, high-value tasks often missed by synthetic methods. Leveraging this dataset, we propose a Dual Mid-Training recipe that decouples spatial grounding from planning, achieving state-of-the-art performance on our proposed WebChainBench and other public GUI benchmarks. Our work provides the data and insights necessary to build and rigorously evaluate the next generation of scalable web agents.
Figures & tables
Feature
WebChain
Mind2Web
WebArena (Env.)
WebLINX
GUIAct(multi)
Scale
Trajectories
31725
2350
N/A
2337
5696
Steps
318k
17155
N/A
100k+
44k
Websites
428
137
4 Domains
155
121
Data Source
Real Websites
✓
✓
×
✓
✓
Table 1: Comparison with Existing Web Interaction Trajectories Datasets
Figure 1: Dataset Overview. A statistical summary of WebChain, including interaction distribution across website categories, top-domain interaction frequencies, device pixel ratio distribution, trajectory complexity, and trajectory duration. These statistics collectively highlight the scale and diversity of WebChain.
Figure 2: Example trajectory and multi-dimensional step information in WebChain. Left: a long-horizon task on Booking.com with key steps along the trajectory. Right: the multi-dimensional step schema, including visual observations, structural semantics, and behavioral annotations.
Metric
Value
Description
Total Trajectories
31,725
Human-verified sessions
Total Steps
317,993
Atomic interactions
Unique Domains
428
E-commerce, OTA, etc.
Avg. Trajectory Length
10.02
Median: 9
Avg. Duration
1.07 min
Time per task
Table 2: High-level statistics of WebChain. The dataset emphasizes long-horizon reasoning with diverse domain coverage.
Models
AC-High [ 12 ]
AC-Low [ 12 ]
GUI-Act-Web [ 2 ]
GUI-Odyssey [ 14 ]
OA-Desktop [ 10 ]
OA-Web [ 10 ]
Overall
Type
GR
SR
Type
GR
SR
Type
GR
SR
Type
GR
SR
Type
GR
SR
Type
GR
SR
Zero Shot
Qwen2.5-VL-3B
49.1
46.1
38.7
61.0
74.6
60.0
56.5
65.9
57.3
37.9
27.1
26.9
58.1
49.0
49.0
51.2
47.0
47.0
50.1
Qwen2.5-VL-7B
68.6
62.3
57.0
87.0
83.0
72.0
86.5
84.2
83.9
55.1
38.1
34.7
84.0
79.3
79.3
80.3
70.3
70.3
70.9
Trained on Other Datasets
GUI-R1-3B
58.0
56.2
46.6
83.7
81.6
64.4
89.9
87.4
76.3
54.8
41.5
41.3
91.9
78.4
78.3
88.6
75.1
75.1
70.5
Table 3: Performance comparison of models across GUI-Act and OmniAct benchmarks. ( AC = AndroidControl , OA = OmniAct )
Action
Short ( <6 )
Medium ( 6–10 )
Long ( >10 )
Click
362
358
365
Type
38
42
35
Table 4: Distribution of WebChainBench samples.
Figure 3: Scaling effects of WebChain subsets (4k, 20k, and Full) on Qwen2.5-VL-3B’s success rate after LCRL post-training.
Figure 4: Study on WCB-S evaluating the effects of Visual Grounding Densification (VGD) and Reasoner Prompting (RP) on spatial grounding performance.
Figure 5: Study on WCB-L evaluating the effect of SGRL Mid-Training on LCRL Post-Training.
GUI-R1-3B
Directly LCRL
+ CoT-SFT
+ SGRL
+Both
WCB-L
0.487
0.603
0.629
0.632
0.658
Table 5: Comparative Analysis of Mid-Training and RL Integration Strategies on WCB-L
Figure 6: Qualitative presentation of the imapact of CoT-SFT mid-training