WebExpert: domain-aware web agents with critic-guided expert experience for high-precision search

A research team led by Yuelin Hu has introduced WebExpert, a novel AI agent designed to tackle the persistent challenge of specialized web searches in domains like finance and biomedicine. Traditional agents often fail when queries drift from their training data or when evidence is noisy. WebExpert addresses this through three key innovations: sentence-level experience retrieval with topic merging, schemalight facet induction that automatically identifies relevant facets like time and policy without hand-written rules, and preference-optimized planning that improves both query formulation and information retrieval.

On benchmark datasets including GAIA, GPQA, and WebWalkerQA, WebExpert demonstrated significant improvements, boosting Answer Exact Match by 1.5 to 3.6 percentage points compared to the strongest existing browsing baselines. The system also reduced the number of page hops required to find answers, making it more efficient. At inference time, a lightweight experience gate helps the agent focus on the most relevant facets while providing fallback options when retrieval confidence is low, creating a more robust browsing experience for complex, domain-specific questions.