Web2BigTable: A Bi-Level Multi-Agent LLM System for Internet-Scale Information Search and Extraction
In a significant advancement in the field of artificial intelligence, researchers have introduced Web2BigTable, a bi-level multi-agent framework designed to enhance web-to-table search capabilities. This innovative system addresses the dual challenges of deep reasoning over specific targets and structured aggregation of information across diverse entities and sources, which have historically posed difficulties for existing systems.
The demand for effective information retrieval has grown tremendously, driven by the need for both breadth and depth in search capabilities. Traditional systems often fall short, particularly when tasked with schema-aligned outputs that require extensive coverage and consistency across multiple entities. Conversely, depth-oriented tasks necessitate coherent reasoning across long and complex search trajectories. Web2BigTable aims to bridge this gap, offering a robust solution for both types of tasks.
Framework Overview
Web2BigTable employs a unique bi-level architecture that separates the orchestration and execution of search tasks. The system comprises an upper-level orchestrator that decomposes complex tasks into manageable sub-problems, enabling lower-level worker agents to tackle these challenges simultaneously. This parallel processing approach not only enhances efficiency but also maximizes the potential for accurate information retrieval.
The framework utilizes a closed-loop process characterized by three key phases: run, verify, and reflect. This iterative cycle allows for continuous improvement in both the decomposition of tasks and the execution of solutions over time. Additionally, Web2BigTable features a persistent, human-readable external memory that facilitates the tracking of progress and learning, enabling self-evolving updates to each agent’s capabilities.
Coordinated Execution
During the execution phase, worker agents collaborate through a shared workspace that enhances visibility into partial findings. This coordination significantly reduces redundant exploration of information, helps reconcile conflicting evidence, and allows agents to adapt to emerging coverage gaps. Such a collaborative environment is crucial for achieving higher accuracy and consistency in results.
Performance Metrics
Web2BigTable has demonstrated remarkable performance in benchmark tests, particularly on the WideSearch dataset. The system achieved an impressive Average@4 Success Rate of 38.50, which is 7.5 times higher than the second-best performing system, which recorded a success rate of only 5.10. In addition, Web2BigTable outperformed the competition with a Row F1 score of 63.53, surpassing the next best result by 25.03, and an Item F1 score of 80.12, exceeding the second best by 14.42.
Moreover, the framework has shown its versatility by generalizing effectively to depth-oriented search tasks on the XBench-DeepSearch dataset, where it achieved an accuracy of 73.0.
Conclusion
Web2BigTable marks a significant leap forward in the realm of information search and extraction, setting a new standard for performance. Its innovative bi-level, multi-agent approach not only enhances the efficiency of data retrieval but also improves the accuracy of results across various tasks. For those interested in exploring the underlying code and capabilities of Web2BigTable, the project is publicly accessible at GitHub.
Related AI Insights
- 3D Layout and Shape Generation from Text Using Diffusion
- Interval Orders & Biorders in Credibility-Limited Belief Revision
- Causal Disentanglement for Accurate Image Quality Assessment
- Binary Spiking Neural Networks Explained with Causal Models
- Ethical Risks of Unilateral Control in Human-AI Relationships
- Open-H-Embodiment: Largest Dataset for Medical Robotics AI
- Personalized Digital Twins for Cognitive Decline Assessment
- Counterfactual Routing to Reduce MoE Model Hallucinations
- Unsupervised Electrofacies & Porosity Analysis in Keta Basin
- Vibe Coding & AI Help-Seeking in Student Programming
