An arXiv preprint reports that Qwen3.5 models fine-tuned with BrowserForge's web-interaction data scored higher than their untuned counterparts on a benchmark of dynamic web tasks. On Online-Mind2Web, the reported end-to-end success rate rose from 25.66% to 33.33% for the 4B model and from 29.33% to 38.00% for the 9B model, gains of 7.67 and 9.33 percentage points, respectively.
On Multimodal-Mind2Web, which reports step-by-step accuracy rather than end-to-end task success, average Pass@1 accuracy was 38.2% for the baseline 4B model and 43.8% after fine-tuning; the corresponding 9B figures were 40.0% and 45.1%. At Pass@4, the averages were 44.1% and 54.4% for the 4B pair and 47.2% and 56.3% for the 9B pair, with the reported improvement holding across Cross-Task, Cross-Website and Cross-Domain splits.
The result is an association within the tested models and benchmarks: the comparison does not isolate open-web sourcing from browser orchestration, task synthesis, verification, filtering or unified reasoning.
How the data was built
BrowserForge combines four pieces: open-web URL sourcing and cleaning, parallel browser-sandbox orchestration, a Proposer-Solver system for task synthesis, and trajectory cleaning with a unified reasoning rewrite. During synthesis, the system can use accessibility-tree information; the trained agent itself acts from screenshots at inference.
A shared work queue lets the cluster manager run up to 300 browser sandboxes concurrently. The Proposer generates three candidate tasks, reflects on them and selects one; the Solver then works through planning, action, reflection and verification.
That process produced 203,238 raw trajectories, one for each distinct website, averaging 8.8 steps apiece, or roughly 1.8 million interaction steps before filtering.
Rule-based and model-based filters retained roughly 30% of the raw interaction steps, leaving about 600,000 verified steps. The final training corpus contained 200,000 Seed-rewritten steps.
The reported gains across tests
In a controlled source comparison using the 4B backbone and a three-epoch training budget, the BrowserForge data scored higher than open-source trajectories on both reported averages: 41.33% versus 32.74% at Pass@1 and 52.59% versus 43.48% at Pass@4.
In a nested data-scaling analysis, Online-Mind2Web success was 25.66% at zero-shot, 27.33% at an intermediate subset and 33.33% on the full set. Multimodal-Mind2Web step accuracy rose from 23.11% to 41.83% and then 54.36% across those same points. The paper describes the trend as approximately power-law shaped and says it had not saturated at the largest reported subset.
In the cleaning ablation, reported Online-Mind2Web success was 33.3% for the full setup, compared with 24.0% without cleaning, 27.0% with rule filtering alone, 29.0% with model judging alone and 31.0% without the unified reasoning rewrite.
The remaining failures matter
In a priority-coded sample of 100 BrowserForge-4B Online-Mind2Web trajectories, 11.0% completed. Click loops accounted for 29.0%, back-navigation loops for 22.0% and scroll-only tails for 20.0%.
The paper identifies a 30-step cap as a partly artificial failure source when a run reaches the limit while making plausible progress.
The reported percentages also come without confidence intervals, statistical tests or repeated-run variability, making the size and stability of the differences difficult to gauge. The benchmarks measure automatic task success or step accuracy, not human outcomes, and the study does not establish how well the method would transfer beyond the tested models, websites and access conditions.
Paper data and sources
Original title: BrowserForge: Scaling Web Episode via Parallel Browser Sandboxes
Authors: Fei Tang, Huawen Shen, Zhiqiong Lu et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-25
DOI: Not available
Original paper · Full text