The paper presents AgentTuning, a method to improve the agent capabilities of large language models (LLMs) without compromising their general abilities. It uses a hybrid instruction-tuning strategy that combines a new dataset, AgentInstruct, with open-source instructions from general domains. The method is applied to the Llama 2 series, resulting in a new model called AgentLM. Evaluation shows that AgentTuning enhances LLMs’ agent abilities without compromising their general abilities, with the AgentLM-70B model performing comparably to GPT-3.5-turbo on unseen agent tasks.
Publication date: 19 Oct 2023
Project Page: https://github.com/THUDM/AgentTuning
Paper: https://arxiv.org/pdf/2310.12823
