SynWeaver: Website-Prior Task and Trajectory Co-Synthesis for Web Agents

2026-08-12

AI Summary

Existing web agents exhibit limited generalization on unseen websites, primarily due to synthetic data that fails to capture full functionality and lacks task authenticity. This work proposes SynWeaver, a novel framework that leverages website prior knowledge to guide task generation. SynWeaver first performs structured exploration to construct a sitemap, extracting page-level and transition-based supervision signals to train a UI-aware model. It then jointly synthesizes semantically consistent and executable task–trajectory pairs, enhanced by a semantic validation and automated repair mechanism that enables iterative co-optimization of tasks and trajectories. Experiments demonstrate that SynWeaver significantly outperforms current methods on WebArena and WebVoyager benchmarks, with the generated data substantially improving agent performance both in-domain and in cross-domain settings.

Abstract

Web agents often struggle to generalize to unseen websites because they lack website-specific supervision. Recent exploration-based data synthesis methods reduce manual annotation, but they still face two key limitations: they often fail to cover the full functionality of a website, and without sufficient website prior knowledge, they tend to propose hallucinated tasks, which in turn limits the diversity and efficiency of downstream trajectory synthesis. We present \textbf{SynWeaver}, a website-prior task-trajectory co-synthesis framework designed to address these challenges. SynWeaver first performs structured website exploration and constructs a website map that covers a broad set of functionally distinct page states and executable interactions on the target website. It then derives page-level and transition-level supervision from this map to train a UI-aware model with website-specific priors, enabling more grounded task proposals. Finally, SynWeaver performs collaborative task-trajectory synthesis, jointly updating the task and execution trajectory when they become inconsistent, and then verifies and repairs the collected results to produce executable, semantically aligned supervision. Experiments on WebArena and WebVoyager demonstrate that SynWeaver consistently outperforms strong synthesis baselines and yields more effective supervision for both in-domain and out-of-domain generalization.

Problem

• web agents
• generalization
• task synthesis
• trajectory synthesis
• website prior

Innovation

• website-prior
• task-trajectory co-synthesis
• structured exploration
• UI-aware model
• executable supervision