核心概念
Large Language Models (LLMs) show potential in simulating user search behaviors, leading to the development of USimAgent for user-centric evaluation.
摘要
1. Introduction
User simulation offers cost-effectiveness and flexibility.
Traditional methods struggle to simulate complex user search behaviors.
Large Language Models (LLMs) show promise in simulating human intelligence.
2. Methodology
USimAgent alternates between generating queries, clicks, and stopping actions.
Reasoning before acting enhances action accuracy.
Query reformulation and click prediction are key components.
3. Experiments
Evaluation on a public user behavior dataset shows USimAgent outperforms existing methods in query generation.
Comparable performance to traditional models in predicting clicks and stopping behaviors.
Ablation study highlights the importance of context information for task completion.
4. Conclusion
USimAgent leverages LLMs effectively for search user simulation.
Future research directions include combining LLMs with broader datasets for improved performance.
統計資料
大規模言語モデル(LLM)は、ユーザー検索行動のシミュレーションに有望性を示しています。
提案されたシミュレータは既存の方法よりもクエリ生成で優れています。
クリックと停止行動の予測において、提案手法は従来のモデルと比較可能な性能を発揮しています。