Reproduce “USimAgent: Large Language Models for Simulating Search Users” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Erhan Zhang, Xingzhu Wang, Peiyuan Gong, Yankai Lin, Jiaxin Mao
USimAgent is a Large Language Model-based framework from Renmin University of China designed to simulate an entire user search session, encompassing query formulation, result clicking, and stopping behaviors. The system generates complete interaction sequences for specific tasks, outperforming traditional methods in query generation by achieving a BLEU score of 0.4630, and showing comparable performance in click prediction and stopping behavior to existing models.
Key insights
- Large Language Models demonstrate strong capabilities in simulating complex, multi-faceted user search behaviors, moving beyond fragmented traditional approaches.
- Maintaining a coherent session context and integrating observations from clicked results is critical for accurate query reformulation and task completion decisions by the LLM.
- The 'Reasoning before Acting' mechanism allows the simulator to mimic cognitive processes like planning and adapting to new information, making simulated behaviors more human-like.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.