Source-linked AI summary
BrowserForge: Scaling Web Episode via Parallel Browser Sandboxes
Fei Tang, Huawen Shen, Zhiqiong Lu, Zhengxi Lu, Pengyuan Lyu, Chengquan Zhang, Weiming Lu, Jun Xiao, Yueting Zhuang, Yongliang Shen
TL;DR
Pure-GUI web agents need large, diverse interaction datasets, but existing sources remain limited in scale or website coverage. BrowserForge addresses this gap by combining open-web sourcing, parallel browser sandboxes, and Proposer–Solver synthesis with verification and cleaning. Its 203,238-trajectory corpus improves performance on both dynamic and static web-agent benchmarks.
Problem
Pure-GUI web agents require many high-quality trajectories across distinct websites, while existing datasets and synthesis pipelines remain limited in scale or website coverage.
Method
BrowserForge sources openly reachable websites, schedules parallel browser sandboxes, and uses a Proposer–Solver loop plus verification and cleaning to produce screenshot-based training trajectories.
Results
203,238 trajectories from distinct websites improve performance across dynamic Online-Mind2Web and static Multimodal-Mind2Web evaluations.
Takeaways & Limitations
Open-web sourcing and broad website coverage are key contributors to the observed improvement, with gains appearing across benchmark protocols and model backbones.
Takeaways & Limitations
Only 11% of sampled BrowserForge-4B Online-Mind2Web trajectories terminate cleanly, with recurring click-loop and back-navigation failures.
Abstract
from arXiv · showhide
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.
1 INTRODUCTION
BrowserForge addresses the need for large, diverse pure-GUI training data by sourcing openly reachable websites and coordinating parallel browser sandboxes. Its 203,238-trajectory corpus improves web-agent benchmark performance across dynamic and static settings.
- Motivation: Pure-GUI agents avoid the fragility and token cost of HTML or accessibility-tree inputs by acting from rendered screenshots.HTML and accessibility structures vary across websites, while a single accessibility tree can serialize to tens of thousands of tokens.
- Motivation: Existing datasets and synthesis pipelines provide limited website diversity because they rely on small site collections, tutorials, or fixed environments.Public datasets contain only a few thousand demonstrations over tens to low hundreds of sites.
- Approach: BrowserForge combines open-web URL sourcing, parallel sandbox orchestration, and a Proposer–Solver loop with verification and cleaning.The pipeline uses page structure only during synthesis; the released agent acts purely from screenshots.
- Results: 203,238 trajectories, each from a distinct website, form a corpus broader than prior trajectory datasets.The framework decouples data scale and diversity from any fixed website list.
- Results: 33.33% Online-Mind2Web success rate and 43.8% Multimodal-Mind2Web average step accuracy demonstrate gains from fine-tuning on the corpus.The reported results are competitive with much larger open-source web agents.
2 RELATED WORK
Prior work synthesizes web-agent trajectories from tutorials, documentation, exploration, or fixed environments, while benchmarks and agents increasingly support visual interaction with rendered pages.
- Automated trajectory synthesis: Automated synthesis methods convert tutorials or other indirect knowledge into executable web-agent demonstrations.AgentTrek and Synatra are described as reusing online tutorials or similar knowledge sources.
- Automated trajectory synthesis: Interaction-based methods generate trajectories through exploration, feedback, relabeling, reverse task synthesis, or documentation-grounded reconstruction.Explorer is reported to produce over 94K trajectories through a bottom-up multi-agent exploration pipeline.
- Web agents, benchmarks, and multimodal GUI models: Web-agent research has shifted from HTML or accessibility-tree interaction toward multimodal agents acting on rendered pages, alongside benchmarks tracking this transition.Mind2Web is cited as offering 2,350 crowdsourced tasks over 137 websites.
3 METHOD
BrowserForge combines open-web sourcing, parallel browser orchestration, Proposer–Solver task synthesis, and trajectory cleaning to build scalable web-agent training data. The pipeline filters and standardizes verified interactions into a broad corpus collected across distinct websites.
- Overview: BrowserForge decomposes into open-web URL sourcing, parallel browser sandbox orchestration, Proposer–Solver task synthesis, and trajectory cleaning.The framework targets scale through parallel sandboxes, diversity through open-web sourcing, and quality through verification and cleaning.
- Open-web URL sourcing and cleaning: Open-web sourcing samples and filters Common Crawl URLs to retain real, reachable, interactive websites rather than a fixed site list.Filtering removes dead, static, near-empty, and otherwise unsuitable pages before sandbox processing.
- Parallel browser sandbox orchestration: Sandbox orchestration maintains shared machines, URL queues, browser pools, and workers so many isolated browser sessions process pages concurrently.The shared queue avoids fixed partitions, allows workers to continue when individual pages stall, and supports adding compute nodes during execution.
- Proposer–Solver task synthesis: The Proposer converts packaged page state into an executable task, while the Solver collects a trajectory through planning, acting, reflection, and verification.The Proposer generates candidate tasks grounded in page elements and excludes tasks requiring registration, login, or payment; the Solver acts on the live page and can recover from intermediate mistakes.
4 EXPERIMENTS
Experiments evaluate BrowserForge on dynamic and static web-agent benchmarks, then isolate the effects of data source, scale, and cleaning. Fine-tuning compact multimodal backbones on BrowserForge data consistently improves performance across protocols and distribution shifts.
- Main Results: BrowserForge data improves both dynamic task success and static step accuracy for the 4B and 9B backbones.Online-Mind2Web and Multimodal-Mind2Web both improve rather than showing gains at only one model or protocol.
- Main Results: 33.33% Online-Mind2Web success is achieved by BrowserForge-4B, up from 25.66%, while BrowserForge-9B reaches 38.00%, up from 29.33%.The corresponding absolute gains are +7.67 and +9.33 percentage points.
- Main Results: 43.8% Pass@1 average step accuracy is reached by BrowserForge-4B on Multimodal-Mind2Web, up from 38.2%, while the 9B model reaches 45.1%, up from 40.0%.The improvements occur on the static benchmark alongside the dynamic success-rate gains.
- Main Results: BrowserForge agents outperform larger open-source models on Online-Mind2Web, with BrowserForge-9B reaching 38.0% versus GUI-Libra-8B at 36.7%.The passage also reports that BrowserForge-9B exceeds cited proprietary reference scores, while BrowserForge uses supervised fine-tuning only.
- Main Results: 54.4% and 56.3% Pass@4 average step accuracy are reached by BrowserForge-4B and 9B, respectively, from backbones at 44.1% and 47.2%.The improvement holds across Cross-Task, Cross-Website, and Cross-Domain splits.
- Data Source and Scale: 41.33% Pass@1 and 52.59% Pass@4 average step accuracy are obtained with BrowserForge data versus 32.74% and 43.48% using matched open-source trajectories.The controlled comparison fixes the Qwen3.5-4B backbone and 3-epoch budget, isolating the training-data source.
5 CONCLUSION
BrowserForge generates large-scale web episodes by running parallel browser sandboxes over openly sourced URLs and cleaning verified trajectories. Its 203,238-trajectory corpus improves benchmark performance when used to fine-tune a compact multimodal model.
- 5 CONCLUSION: BrowserForge decouples web-agent data scale and diversity from fixed website lists by using parallel sandboxes over openly sourced URLs.The framework combines open-web sourcing, shared-queue sandbox scheduling, and Proposer–Solver task and trajectory generation.
- 5 CONCLUSION: 203,238 trajectories, each from a distinct website, form a corpus larger and broader than prior trajectory datasets.
- 5 CONCLUSION: Fine-tuning a compact multimodal model raises Online-Mind2Web success rate from 25.66% to 33.33% and improves Multimodal-Mind2Web step accuracy.Controlled analyses trace the gain to open-web data and website diversity rather than the training recipe alone.
A APPENDIX
The appendix will provide additional implementation details, dataset examples, and qualitative trajectories.
- A APPENDIX: Additional implementation details will be provided in the appendix.
- A APPENDIX: The appendix will include dataset examples.
- A APPENDIX: The appendix will provide qualitative trajectories.
A.1 TRAINING AND INFERENCE SYSTEM PROMPT
The training and inference system prompt asks a browser GUI agent to use the instruction, current screenshot, and action history to predict one next action. It standardizes output formatting and coordinate conventions across training and evaluation.
- A.1 TRAINING AND INFERENCE SYSTEM PROMPT: The agent predicts a single next action from the instruction, current screenshot, and previous-action history.
- A.1 TRAINING AND INFERENCE SYSTEM PROMPT: Coordinates use integer values normalized to the [0, 1000] range regardless of image resolution.
- A.1 TRAINING AND INFERENCE SYSTEM PROMPT: The unified action space includes operations such as clicking, typing, selecting, and scrolling.
- A.1 TRAINING AND INFERENCE SYSTEM PROMPT: The [EXIT] message terminates the task when the requested answer was already provided in a previous step.
- A.1 TRAINING AND INFERENCE SYSTEM PROMPT: The model must output Thought, Description, and Action as exactly three lines in that order.