SynWeaver framework helps web agents tackle unseen websites

Web agents—AI systems that navigate websites to complete tasks—often stumble on sites they haven't seen during training. Existing exploration-based data synthesis methods reduce manual labeling but fail to cover full site functionality and can hallucinate tasks. SynWeaver, created by Ruitao Wang, Yuwen Hao, and Menglin Yang, tackles this by first performing structured website exploration to build a 'website map' that captures distinct page states and executable interactions. From this map, it derives page-level and transition-level supervision to train a UI-aware model with website-specific priors, making task proposals more grounded in reality.

The framework then runs collaborative task-trajectory synthesis, where the task and its execution path are updated together if they become inconsistent, followed by verification and repair to produce executable, semantically aligned supervision. Compared against strong synthesis baselines on WebArena and WebVoyager benchmarks, SynWeaver consistently generates more effective supervision, improving both in-domain and out-of-domain generalization. For builders of automated web agents, this means fewer manual annotations and better performance on unseen sites—a meaningful step toward agents that can operate across the open web rather than only in curated environments.