# AgentTuning: Enabling Generalized Agent Abilities for LLMs

**Authors:** Aohan Zeng, Mingdao Liu, Rui Lu, Bowen Wang, Xiao Liu, Yuxiao Dong, Jie Tang  
**arXiv:** [2310.12823](https://arxiv.org/abs/2310.12823)  
**Date:** 2023-10-19  
**Summary page:** https://www.summarizepaper.com/en/paper/2310.12823v2/

## Key points

- Researchers introduce AgentTuning to enhance agent capabilities of large language models (LLMs) while maintaining general abilities
- Importance of fine-grained prompting methods and robust LLMs for satisfactory performance in tasks where LLMs act as central controllers
- AgentTuning is a simple yet effective approach to improving LLMs' agent abilities without compromising overall functionality
- Creation of lightweight instruction-tuning dataset called AgentInstruct containing high-quality interaction trajectories
- Successful instruction-tuning of Llama 2 series to create AgentLM by combining it with open-source instructions using a hybrid strategy
- Evaluation shows significant boost in LLMs' agent capabilities without sacrificing general performance
- Resulting models (7B, 13B, and 70B variants) exhibit comparable performance to commercial models on unseen agent tasks
- Generalized agent capabilities achieved through AgentTuning contribute valuable insights into advancing LLM technology for real-world applications

## Summary

, , , , 

In the study "AgentTuning: Enabling Generalized Agent Abilities for LLMs," researchers Aohan Zeng, Mingdao Liu, Rui Lu, Bowen Wang, Xiao Liu, Yuxiao Dong, and Jie Tang introduce a novel method called AgentTuning to enhance the agent capabilities of large language models (LLMs) while preserving their general abilities. The team highlights the importance of fine-grained prompting methods and robust LLMs for satisfactory performance in tasks where LLMs act as central controllers. To address this gap, they present AgentTuning as a simple yet effective approach to improving LLMs' agent abilities without compromising their overall functionality. The researchers create a lightweight instruction-tuning dataset called AgentInstruct containing high-quality interaction trajectories. By combining it with open-source instructions from general domains using a hybrid instruction-tuning strategy, they successfully instruction-tune the Llama 2 series to create AgentLM. Through evaluations, they demonstrate that AgentTuning significantly boosts LLMs' agent capabilities without sacrificing their general performance. The resulting models - including 7B, 13B, and 70B variants - exhibit comparable performance to commercial models on unseen agent tasks. These findings showcase the generalized agent capabilities achieved through AgentTuning and contribute valuable insights into advancing LLM technology for real-world applications. Furthermore, the researchers make their datasets and trained models openly available on GitHub at https://github.com/THUDM/AgentTuning to provide accessible alternatives to commercial LLMs for handling diverse agent tasks effectively.