WebExpert: Domain-Aware Web Agents with Critic-Guided Expert Experience for High-Precision Search
Summary: arXiv:2604.06177v1 Announce Type: cross
In today’s rapidly evolving digital landscape, specialized web tasks in sectors such as finance, biomedicine, and pharmaceuticals face significant challenges. The complexities are primarily due to the absence of robust domain priors, which lead to query drift, noisy evidence, and brittle reasoning. To address these issues, researchers have unveiled WebExpert, a cutting-edge domain-aware web agent designed to enhance the precision of web search.
Overview of WebExpert
WebExpert is implemented as an end-to-end solution that incorporates several innovative features aimed at improving the effectiveness of web searches. The core functionalities of WebExpert include:
- Sentence-Level Experience Retrieval: This feature integrates topic merging and rule distillation to refine the retrieval process, enhancing the quality of results.
- Schemalight Facet Induction: Unlike traditional methods that rely on static hand-written lexicons, WebExpert utilizes weak supervision to bootstrap facets such as time, region, policy, and industry.
- Preference-Optimized Planning: By employing pairwise preference learning and a coverage-aware objective, WebExpert improves both query planning and retrieval processes.
How WebExpert Works
At the inference stage, WebExpert utilizes a lightweight experience gate. This gate biases the decoding process toward active facets while providing a fallback mechanism in scenarios of low retrieval confidence. This approach ensures that users receive the most relevant information even when initial searches yield limited results.
Performance and Results
The effectiveness of WebExpert has been evaluated using several benchmarks, including GAIA, GPQA, HLE, and WebWalkerQA. The results show that WebExpert significantly outperforms traditional browsing baselines, with improvements in Answer Exact Match (EM) by 1.5 to 3.6 percentage points. Furthermore, it reduces the number of page hops required to find accurate answers, thereby enhancing the overall user experience.
Analysis and Future Directions
In-depth analysis of WebExpert reveals consistent gains across various components, including retrieval, topic merging, facet induction, and preference-aware training. These findings suggest that the integration of domain knowledge and user preferences can lead to more effective web search solutions. As the field of AI continues to evolve, further research will focus on refining these techniques and exploring new applications across different domains.
Conclusion
WebExpert represents a significant advancement in the development of domain-aware web agents. By combining innovative retrieval techniques with user-centric planning, it addresses the challenges posed by specialized web tasks. As the demand for high-precision search capabilities grows, solutions like WebExpert will play an essential role in shaping the future of web search technology.