This is a page that curates AI-related papers published worldwide. All content here is summarized using Google Gemini and operated on a non-profit basis. Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.
This paper focuses on solving complex and open problems by leveraging the web-based information seeking (IS) capabilities of large-scale language model (LLM)-based agents. Existing information-based approaches generate questions after collecting web data, which can cause mismatches between information structure and inference structure, and between questions and answers. To address this, this paper proposes WebShaper, a data synthesis framework that systematically formulates IS tasks based on set theory. WebShaper precisely controls the inference structure centered on the concept of knowledge projection (KP), and generates datasets through a multi-stage expansion process after seed task generation. The agent-based expander uses search and verification tools based on the formulation to expand the current formulated questions to make them more complex. By training the model with the synthesized dataset, WebShaper achieves state-of-the-art performance among open-source IS agents on the GAIA and WebWalkerQA benchmarks.
Takeaways, Limitations
•
Takeaways:
◦
We present WebShaper, a novel data synthesis framework for improving the performance of web-based information seeking agents.
◦
Solving the mismatch problem between information structure and inference structure through a set theory-based formalism.
◦
Precise control of inference structures using the concept of knowledge projection (KP).
◦
Achieving state-of-the-art performance on GAIA and WebWalkerQA benchmarks.
•
Limitations:
◦
WebShaper's performance improvements may be limited to specific benchmarks.
◦
Further validation of generalization performance on synthetic datasets is needed.
◦
Further research is needed on its applicability to complex and diverse information seeking tasks in the real world.
