Large language models (LLMs) have been successful in various natural language processing (NLP) tasks and have shown potential in achieving some aspects of artificial general intelligence. Recent research suggests that supplementing LLMs with external tools can enhance their problem-solving capabilities, similar to how human intelligence has evolved. However, the availability of suitable tools is critical for the effectiveness of these procedures.
In this study, researchers from Google Deepmind, Princeton University, and Stanford University propose a system called LLMs As Tool Makers (LATM), inspired by the importance of tool-making in human development. LATM allows LLMs to create their own reusable tools to tackle new tasks. The process involves two essential phases: tool creation and tool application.
During the tool creation phase, an LLM called the tool builder develops tools, implemented as Python functions, specifically for a particular task. In the tool application phase, another LLM known as the tool user, who may be the same as the tool builder, utilizes these tools to address new requests. LATM ensures that the most qualified LLM is assigned to each step, thereby optimizing the workflow.
The two-stage design of LATM allows for the use of a powerful, resource-intensive model like GPT-4 for tool creation. For the tool application, a lightweight and affordable model such as GPT-3.5 Turbo can be employed, significantly reducing computing costs without compromising problem-solving abilities. The tool-making process only needs to be carried out once for a specific capability, and the produced tools can be applied to multiple task instances.
This approach provides a scalable and cost-effective solution for handling challenging problems. For instance, if a user asks the LLM to schedule a meeting, a lightweight model might struggle with complex arithmetic reasoning, while a more powerful model can provide accurate answers at higher inference costs. By combining the strengths of these models, LATM overcomes these obstacles.
LATM can also be used in various applications, including solving well-known games like Sudoku, parsing and analyzing online articles, and creating specialized routing plans. Additionally, LATM introduces a dispatcher, another lightweight LLM, which determines whether an incoming problem can be solved using existing tools or if a new tool needs to be developed. This dynamic architecture enables real-time tool creation and usage.
The trials conducted by the researchers demonstrate the effectiveness of LATM on challenging problems and complex thinking tasks. The results indicate that LATM can perform as well as more resource-intensive models while being more cost-effective. This unique approach to LLMs, inspired by human tool-making, opens up exciting possibilities for a society that relies on LLM-generated tools.
To learn more about LATM, you can refer to the paper and GitHub link provided. Stay updated with the latest AI research news, projects, and more by joining our ML subreddit, Discord channel, and email newsletter. If you have any questions or feedback, feel free to reach out to us at Asif@marktechpost.com.
About the Author:
Aneesh Tickoo is a consulting intern at MarktechPost, currently pursuing an undergraduate degree in Data Science and Artificial Intelligence from the Indian Institute of Technology (IIT), Bhilai. He is passionate about harnessing the power of machine learning and focuses on image processing research. Aneesh enjoys collaborating on interesting projects and connecting with people in the field.