AgentTuning: Enabling Generalized Agent Abilities for LLMs
(2310.12823)Abstract
Open LLMs with great performance in various tasks have significantly advanced the development of LLMs. However, they are far inferior to commercial models such as ChatGPT and GPT-4 when acting as agents to tackle complex tasks in the real world. These agent tasks employ LLMs as the central controller responsible for planning, memorization, and tool utilization, necessitating both fine-grained prompting methods and robust LLMs to achieve satisfactory performance. Though many prompting methods have been proposed to complete particular agent tasks, there is lack of research focusing on improving the agent capabilities of LLMs themselves without compromising their general abilities. In this work, we present AgentTuning, a simple and general method to enhance the agent abilities of LLMs while maintaining their general LLM capabilities. We construct AgentInstruct, a lightweight instruction-tuning dataset containing high-quality interaction trajectories. We employ a hybrid instruction-tuning strategy by combining AgentInstruct with open-source instructions from general domains. AgentTuning is used to instruction-tune the Llama 2 series, resulting in AgentLM. Our evaluations show that AgentTuning enables LLMs' agent capabilities without compromising general abilities. The AgentLM-70B is comparable to GPT-3.5-turbo on unseen agent tasks, demonstrating generalized agent capabilities. We open source the AgentInstruct and AgentLM-7B, 13B, and 70B models at https://github.com/THUDM/AgentTuning, serving open and powerful alternatives to commercial LLMs for agent tasks.
Overview
-
The paper introduces AgentTuning, a novel approach to enhance the agent capabilities of LLMs through a lightweight dataset (AgentInstruct) and a hybrid instruction-tuning strategy.
-
The research was evaluated on the Llama 2 series, resulting in the creation of AgentLM, which demonstrated competitive performance with commercial models like GPT-3.5-turbo on sophisticated agent tasks.
-
AgentInstruct and the resulting models (AgentLM-7B, 13B, 70B) are open-source, providing robust alternatives to commercial LLMs for various agent tasks, maintaining general abilities without compromise.
Overview of AgentTuning: Enabling Generalized Agent Abilities for LLMs
AgentTuning: Enabling Generalized Agent Abilities for LLMs presents a novel approach designed to enhance the agent capabilities of LLMs. This work introduces two key components: AgentInstruct, a lightweight instruction-tuning dataset, and a hybrid instruction-tuning strategy combining agent-specific data with general-domain instructions. The primary goal of this research is to improve the agent capabilities of LLMs without sacrificing their general utility. The proposed method was evaluated on the Llama 2 series, resulting in the creation of AgentLM, which demonstrated competitive performance with commercial models like GPT-3.5-turbo on sophisticated agent tasks.
Key Contributions
- AgentInstruct Dataset: A high-quality instruction-tuning dataset constructed to enhance generalized agent abilities. This dataset consists of 1,866 interaction trajectories across varying tasks, such as ALFWorld, WebShop, and Mind2Web.
- Hybrid Instruction-Tuning Strategy: A methodology combining AgentInstruct with open-source instructions from general domains. This approach allows the Llama 2 series to be fine-tuned into AgentLM models, preserving their general abilities while boosting their effectiveness in agent-specific tasks.
- Evaluation and Results: AgentTuning significantly enhances LLMs' agent capabilities, with AgentLM-70B matching GPT-3.5-turbo on unseen agent tasks. The open-source dataset and models (AgentLM-7B, 13B, and 70B) have been released, providing robust alternatives to commercial LLMs for agent tasks.
Evaluation and Performance
Performance Evaluation:
- Held-in and Held-out Tasks: AgentLM demonstrated strong performance on both sets, with AgentLM-70B achieving results comparable to GPT-3.5-turbo.
- General Tasks: AgentLM maintained its capabilities in traditional NLP benchmarks such as MMLU, GSM8K, and HumanEval, indicating no compromise in general abilities.
- Error Reduction: An analysis showed a significant reduction in elementary mistakes such as formatting errors and duplicated generations in AgentLM, illustrating the efficacy of the tuning process in enabling agent capabilities.
Detailed Contributions
AgentInstruct Construction:
- Diverse Tasks: The dataset spans a variety of agent tasks from practical real-world scenarios, ensuring robustness and generalization.
- Trajectory Interaction & Filtering: High-quality interaction trajectories were generated using GPT-4, followed by automated filtering based on task-specific rewards to ensure data quality.
Hybrid Instruction-Tuning Strategy:
- Dataset Mixture: By blending agent-specific data with general-domain instructions, the model balances specialized abilities and general performance.
- Empirical Validation: Extensive evaluations verified that a mixture of 0.2 ratio of agent data to general data provided optimal results, particularly in the ability to generalize across diverse tasks.
AgentLM's Practical Implications:
- Open Source Release: The provision of the AgentInstruct dataset and AgentLM models supports further research and advances in LLM applications.
- Performance Benchmarking: Established a new benchmark for open-source models in agent tasks, significantly narrowing the gap between open-source and commercial models.
Implications and Future Directions
Theoretical Implications:
- The study underscores the importance of integrating domain-specific tasks with general capabilities to produce versatile LLMs.
- Proposes a scalable method for improving agent abilities that could be adapted to various LLM architectures.
Practical Implications:
- Enhances the utility of open-source LLMs, promoting broader accessibility and application.
- Provides a foundation for developing more sophisticated autonomous agents capable of complex real-world interaction tasks.
Future Developments:
- Expanding the dataset to include more diverse scenarios and tasks, potentially improving the robustness and versatility of the models.
- Further fine-tuning and testing with larger datasets to explore the limits of the hybrid instruction-tuning strategy.
Conclusion
AgentTuning: Enabling Generalized Agent Abilities for LLMs offers a significant contribution to the field by providing a method to enhance the agent capabilities of LLMs without compromising their general abilities. By introducing AgentInstruct and a hybrid instruction-tuning strategy, the research demonstrates the feasibility of achieving robust agent performance comparable to commercial models using open-source LLMs. This work not only bridges the gap between open-source and commercial models but also facilitates further research and development in the domain of autonomous agents.