BrowserForge: Scaling Web Episode via Parallel Browser Sandboxes

Fei Tang; Huawen Shen; Zhiqiong Lu; Zhengxi Lu; Pengyuan Lyu; Chengquan Zhang; Weiming Lu; Jun Xiao

cs.CLAug 25, 2026

BrowserForge: Scaling Web Episode via Parallel Browser Sandboxes

Authors: Fei Tang, Huawen Shen, Zhiqiong Lu, Zhengxi Lu, Pengyuan Lyu, Chengquan Zhang, Weiming Lu, Jun Xiao, +2 more

Organizations: Zhejiang University · LLM Department, Tencent

Abstract

Web agents that act from rendered pixels avoid the fragility and heavy token cost of reading a page's HTML or accessibility tree, but training them depends on large amounts of high-quality interaction trajectories, and how to produce such data at scale remains an open problem. Public datasets typically contain only a few thousand trajectories drawn from a fixed and narrow set of websites, and even recent automated synthesis pipelines stay bound to predefined site lists or tutorial sources, so the number of distinct websites the agent ever sees barely grows. We present BrowserForge, a framework that generates web interaction data at scale by driving many browser sandboxes in parallel over the open web. BrowserForge couples three components: an open-web sourcing stage that exposes the agent to hundreds of thousands of real, openly reachable websites; a sandbox cluster manager that schedules hundreds of concurrent browsers with high utilization; and a Proposer-Solver dual-agent loop that turns a raw page into an executable task and then collects a verified trajectory for it. A rule-plus-model cleaning pipeline removes failed runs and rewrites the surviving reasoning into a single unified chain-of-thought style. Page structure such as the accessibility tree is used only as a synthesis-time signal; the agent we train and release acts purely from the screenshot. The resulting corpus contains 203,238 trajectories, each collected from a distinct website, larger and more diverse than prior trajectory datasets. Fine-tuning a compact multimodal model on this corpus raises its success rate on the live Online-Mind2Web from 25.66% to 33.33% and consistently improves step accuracy on the static Multimodal-Mind2Web, with the gain growing as the corpus scales. Controlled analyses further confirm that open-web sourcing and broad website coverage are key contributors to the observed improvement.

Figures & tables

Figure 1: The BrowserForge synthesis framework. A sandbox cluster manager (left) maintains a machine list, a URL work queue, a pool of browser sandboxes, and worker threads. For each page, the Proposer (middle) generates a candidate task in two stages and selects the most executable one; the Solver (right), built on an open browser-control agent, executes the task through a plan–act–reflect–verify loop and writes the trajectory to storage. Verified trajectories are cleaned and formatted into the final training corpus.

Action	Arguments	Description
click	coordinate	Click a point on the screen.
type	coordinate, text	Focus an input field and type text.
select	coordinate, value	Choose a value from a dropdown / combobox.
scroll	direction, amount	Scroll the view up/down/left/right by a fraction of the viewport.
key	key	Press a single key (e.g. Enter, Tab, Escape).
wait	seconds	Wait for the page to load or settle.

Table 1: The Solver’s unified action space. Screen positions are expressed as normalized integer coordinates in

[0,1000]

, independent of the rendered resolution. At each step the Solver emits exactly one action from this set; the finish action terminates the trajectory.

Dataset	Scale		Diversity	Reason
Dataset	#Trajectories	#Websites	Diversity	Reason
Mind2Web ( Deng et al., 2023 )	1,009	137	Fixed	$\times$
WebLINX ( Lù et al., 2024 )	969	155	Fixed	$\times$
GUIAct ( Chen et al., 2024 )	2,482	121	Fixed	$\times$
OpenWebVoyager ( He et al., 2024b )	1,165	48	Fixed	$\checkmark$
NNetNav ( Murty et al., 2024 )	10,272	20	Fixed	$\checkmark$

Table 2: Comparison of BrowserForge with existing web-agent trajectory datasets. “Reason” marks whether task synthesis includes an explicit reasoning/verification step. Each BrowserForge trajectory is collected from a distinct website, so its trajectory and website counts coincide; it is the largest in trajectory count and the broadest in website coverage, sourcing sites from the open web rather than a fixed list.

Model	Domain	Agent Traj.	SR (%)
Proprietary Models
Browser-Use	Web Agent	N/A	30.0
Gemini-CUA	Computer-Use Agent	N/A	69.0
Navigator	Web Agent	N/A	78.7
Operator	Computer-Use Agent	Undisc.	61.3
GPT-5 + UGround-v1-7B	Planner & Grounder	N/A	33.3

Table 3: Success rate (SR, %) on Online-Mind2Web, scored using the WebJudge-7B automatic judge. Agent Traj. denotes the reported number of agent-interaction trajectories used for specialized training. “Undisc.” indicates that the training scale is not publicly disclosed, while “–” indicates that no specialized agent trajectories are used. Proprietary models are included for reference only. Bold denotes the best result among the compared base models, existing methods, and our models; underlining denotes the best prior result. BrowserForge-4B/9B are fine-tuned from Qwen3.5-4B/9B on the BrowserForge corpus, with absolute gains over the corresponding backbones shown in green.

Proprietary Models with SeeAct-V Framework
Model	Cross-Task		Cross-Website		Cross-Domain		Average
Model	Pass@1	Pass@4	Pass@1	Pass@4	Pass@1	Pass@4	Pass@1	Pass@4
GPT-4o + UGround-v1-7B	35.7	38.9	33.9	37.6	39.1	42.2	36.2	39.6
GPT-4.1 + UGround-v1-7B	41.1	44.8	36.2	39.7	43.0	46.4	40.1	43.6
GPT-5-mini + UGround-v1-7B	44.2	48.0	40.4	44.2	45.8	48.1	43.5	46.7
GPT-5 + UGround-v1-7B	47.7	51.7	45.0	47.6	48.2	51.6	47.0	50.3

Table 4: Step accuracy (%) on Multimodal-Mind2Web. Proprietary models are shown for reference. Bold : best; underline : second best among open-source methods. BrowserForge-4B/9B are fine-tuned from Qwen3.5-4B/9B on the BrowserForge corpus and improve over their respective backbones on every split.

Training data	Cross-Task		Cross-Website		Cross-Domain		Average
Training data	Pass@1	Pass@4	Pass@1	Pass@4	Pass@1	Pass@4	Pass@1	Pass@4
Open-source trajectories	36.45	45.41	28.85	41.02	32.93	44.01	32.74	43.48
BrowserForge (Ours)	44.88	54.74	37.88	50.83	41.22	52.20	41.33	52.59

Table 5: Controlled comparison of the training-data source on Multimodal-Mind2Web step accuracy (%). Holding the Qwen3.5-4B backbone and the

3

-epoch training budget fixed, we fine-tune on existing open-source web trajectories versus the BrowserForge corpus. Read across the two rows: the gap isolates the effect of the data source rather than the architecture or optimizer.

Figure 2: (a) Data scaling. Overall result versus the number of BrowserForge training samples (log scale), fine-tuning Qwen3.5-4B for

3

epochs under a fixed budget, on both Multimodal-Mind2Web (step accuracy) and Online-Mind2Web (success rate). The

0

-sample point is the zero-shot base model; solid curves are power-law fits over the measured points and dashed segments extrapolate beyond the largest run. (b) Cleaning ablation. Online-Mind2Web success rate for each variant of the cleaning pipeline (rule-based filter, model-based judge, and Seed-rewritten unified chain-of-thought), with the full system highlighted.

Outcome	Description	%
Task completed	Terminates with Finish / Exit or a final answer	11.0
Click loop	Same coordinate clicked $\geq 4$ times with no state change	29.0
Back loop	$\geq 3$ GoBack actions (often paired with scrolling)	22.0
Scroll-only tail	Run ends with $\geq 5$ consecutive Scroll actions	20.0
Access blocked	Reasoning mentions access denied / CAPTCHA / bot wall	9.0
Step limit	Exhausts the $30$ -step budget without an above pattern	8.0

Table 6: Outcome distribution over

100

BrowserForge-4B trajectories sampled from Online-Mind2Web. The top row is the only success category; the remaining rows are mutually exclusive failure modes, assigned by a priority rule and ordered by frequency. The three dominant failures, namely repeated clicking, repeated back-navigation, and trailing scrolls, are all forms of unproductive repetition that together account for

71\%

of trajectories.

Explore similar work

Mar 5, 2026

WebChain: A Large-Scale Human-Annotated Dataset of Real-World Web Interaction Traces

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.

Sicheng Fan, Rui Wan, Yifei Leng +4

Fudan University · IMean AI · WebAgentLab

Web Agents Web Agent Benchmarks

Ask PDF

Apr 9, 2026

MolmoWeb: Open Visual Web Agent and Open Data for the Open Web

Web agents--autonomous systems that navigate and execute tasks on the web on behalf of users--have the potential to transform how people interact with the digital world. However, the most capable web agents today rely on proprietary models with undisclosed training data and recipes, limiting scientific understanding, reproducibility, and community-driven progress. We believe agents for the open web should be built in the open. To this end, we introduce (1) MolmoWebMix, a large and diverse mixture of browser task demonstrations and web-GUI perception data and (2) MolmoWeb, a family of fully open multimodal web agents. Specifically, MolmoWebMix combines over 100K synthetic task trajectories from multiple complementary generation pipelines with 30K+ human demonstrations, atomic web-skill trajectories, and GUI perception data, including referring expression grounding and screenshot question answering. MolmoWeb agents operate as instruction-conditioned visual-language action policies: given a task instruction and a webpage screenshot, they predict the next browser action, requiring no access to HTML, accessibility trees, or specialized APIs. Available in 4B and 8B size, on browser-use benchmarks like WebVoyager, Online-Mind2Web, and DeepShop, MolmoWeb agents achieve state-of-the-art results outperforming similar scale open-weight-only models such as Fara-7B, UI-Tars-1.5-7B, and Holo1-7B. MolmoWeb-8B also surpasses set-of-marks (SoM) agents built on much larger closed frontier models like GPT-4o. We further demonstrate consistent gains through test-time scaling via parallel rollouts with best-of-N selection, achieving 94.7% and 60.5% pass@4 (compared to 78.2% and 35.3% pass@1) on WebVoyager and Online-Mind2Web respectively. We will release model checkpoints, training data, code, and a unified evaluation harness to enable reproducibility and accelerate open research on web agents.

Tanmay Gupta, Piper Wolters, Zixian Ma +13

Allen Institute for AI · University of Washington · UNC-Chapel Hill

Web Agents LLM Agent Training Tool-Using Vision-Language Agents

Ask PDF

Apr 13, 2026

WebForge: Breaking the Realism-Reproducibility-Scalability Trilemma in Browser Agent Benchmark

Existing browser agent benchmarks face a fundamental trilemma: real-website benchmarks lack reproducibility due to content drift, controlled environments sacrifice realism by omitting real-web noise, and both require costly manual curation that limits scalability. We present WebForge, the first fully automated framework that resolves this trilemma through a four-agent pipeline -- Plan, Generate, Refine, and Validate -- that produces interactive, self-contained web environments end-to-end without human annotation. A seven-dimensional difficulty control framework structures task design along navigation depth, visual complexity, reasoning difficulty, and more, enabling systematic capability profiling beyond single aggregate scores. Using WebForge, we construct WebForge-Bench, a benchmark of 934 tasks spanning 7 domains and 3 difficulty levels. Multi-model experiments show that difficulty stratification effectively differentiates model capabilities, while cross-domain analysis exposes capability biases invisible to aggregate metrics. Together, these results confirm that multi-dimensional evaluation reveals distinct capability profiles that a single aggregate score cannot capture. Code and benchmark are publicly available at https://github.com/yuandaxia2001/WebForge.

Peng Yuan, Yuyang Yin, Yuxuan Cai +1

Tencent BAC · Tsinghua University

Web Agent Benchmarks Benchmark Design AI Agent Evaluation Synthetic Benchmark Generation

Ask PDF