# Proactive Agent: Shifting LLM Agents from Reactive Responses to Active Assistance

**Authors:** Yaxi Lu, Shenzhi Yang, Cheng Qian, Guirong Chen, Qinyu Luo, Yesai Wu, Huadong Wang, Xin Cong, Zhong Zhang, Yankai Lin, Weiwen Liu, Yasheng Wang, Zhiyuan Liu, Fangming Liu, Maosong Sun  
**arXiv:** [2410.12361](https://arxiv.org/abs/2410.12361)  
**Date:** 2024-10-16  
**Summary page:** https://www.summarizepaper.com/en/paper/2410.12361v3/

## Key points

- Authors propose a novel data-driven approach to develop proactive agents capable of anticipating and initiating tasks without explicit human instructions
- Real-world human activities are collected to generate proactive task predictions and labeled as accepted or rejected by human annotators
- A reward model is trained to evaluate the proactiveness of Large Language Model (LLM) agents automatically
- Comprehensive data generation pipeline creates a diverse dataset called ProactiveBench containing 6,790 events
- Fine-tuning models with this dataset significantly enhances the proactiveness of LLM agents, achieving an F1-Score of 66.47% in proactively offering assistance
- Research paves the way for future advancements in human-agent collaboration by enabling agents to anticipate and initiate tasks autonomously with foresight

## Summary

In their paper titled "Proactive Agent: Shifting LLM Agents from Reactive Responses to Active Assistance," authors Yaxi Lu, Shenzhi Yang, Cheng Qian, Guirong Chen, Qinyu Luo, Yesai Wu, Huadong Wang, Xin Cong, Zhong Zhang, Yankai Lin, Weiwen Liu, Yasheng Wang, Zhiyuan Liu, Fangming Liu and Maosong Sun address the limitations of reactive agent systems by proposing a novel data-driven approach to develop proactive agents capable of anticipating and initiating tasks without explicit human instructions. The authors collect real-world human activities to generate proactive task predictions and label these predictions as accepted or rejected by human annotators. They then train a reward model that simulates human judgment to evaluate the proactiveness of Large Language Model (LLM) agents automatically. Furthermore, the authors develop a comprehensive data generation pipeline to create a diverse dataset called ProactiveBench containing 6,790 events. By fine-tuning models with this dataset, they demonstrate that the proposed approach significantly enhances the proactiveness of LLM agents. Experimental results show that their fine-tuned model achieves an impressive F1-Score of 66.47% in proactively offering assistance, surpassing both open-source and close-source models. These findings underscore the potential of the authors' method in creating more proactive and effective agent systems. Their research paves the way for future advancements in human-agent collaboration by enabling agents to anticipate and initiate tasks autonomously with foresight. The study contributes valuable insights into enhancing the capabilities of agent systems powered by large language models for improved performance in complex scenarios requiring proactive decision-making.