Public trajectories for front-end agent training have been scarce. WebUniverse releases 15K executable environments and verified trajectories, and fine-tuning on them lifted a 35B-A3B model's Vision2Web score from 51.7 to 68.3.
Key takeaways · AI-distilled
WebUniverse builds verifiable front-end development tasks from open-source repositories and real web pages, then keeps AI agentAn AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.Full definition → trajectories only after checking the resulting apps against functional or visual requirements.
The release includes 15K executable task environments (WebUniverse-Env) and 15K matching development trajectories with verification records and scores (WebUniverse-Mix).
fine-tuningTaking a trained model and training it a bit more on your own examples so it gets better at one specific job.Full definition → Qwen3.5-35B-A3B on the trajectories lifted WebGen-Bench accuracy from 34.5% to 63.9%, Vision2Web from 51.7 to 68.3, and Web-Bench pass@2 from 12.2% to 23.1%, per the abstract.
Ablations reported in the abstract show text and visual trajectories add complementary gains, and results kept improving as training data grew from 5K to 15K trajectories.
Terms in this piece · Glossary
fine-tuning — Taking a trained model and training it a bit more on your own examples so it gets better at one specific job.
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
