Improving Web Agent Performance via Co-Synthesis

Summary & Methodology Analysis

SynWeaver tackles the inability of web agents to handle unseen sites by creating website-specific supervision. The process begins with a DFS-based crawler that maps the target website to identify distinct UI states and transitions. This map provides the necessary data to fine-tune UI-aware models, which are neural networks adjusted to understand and interact with interface elements. By leveraging this structured map, the system derives page-level and transition-level supervision, transforming raw website interactions into actionable data for the agents. The core innovation lies in the collaborative task-trajectory synthesis, where the system updates task intents and execution prefixes simultaneously using a teacher model to ensure consistency. This dual-optimization strategy reduces hallucinated task proposals that typically plague agents when navigating unfamiliar environments. The pipeline further enhances data quality through a reconstruction loop that triggers programmatic checks, specifically verifying termination, validity, and consistency, whenever a task-trajectory pair fails. This allows for the iterative expansion of the dataset by resampling transitions and prompting the model to generate a wider variety of tasks. Regarding performance, the method shows robust results on standard benchmarks. On WebArena, the approach reaches a 19.91 success rate using Qwen3-VL-8B-Instruct and 14.16 using InternVL3-8B. On WebVoyager, it achieves a 27.06 success rate, representing a 4.64 point lead over the strongest existing baseline. However, the system faces practical limitations in production scenarios. Since it relies on automated crawling, it is currently vulnerable to anti-bot measures like CAPTCHAs, rate limiting, and dynamic authentication. Furthermore, the student models remain constrained by the reasoning and interaction performance of the proprietary teacher models employed during synthesis.