Large-scale language models (LLMs) excel at a wide range of NLP tasks, showing encouraging evidence that they can achieve some of the capabilities of artificial general intelligence. Similar to the evolution of human intelligence, recent research has also revealed the potential to supplement LLM with external tools to significantly improve problem-solving ability and efficiency. However, the availability of suitable tools is a major factor in determining how applicable procedures using these tools are. According to the lessons learned from these milestones, people’s ability to create tools to solve new problems was an important turning point in human development.
In this study, researchers from Google Deepmind, Princeton University, and Stanford University apply this evolutionary concept, motivated by the importance of tool-making to humans, to the field of LLM. The system they propose is called LLM As Tool Makers (LATM) and will allow LLMs to create reusable tools to take on new responsibilities. Their strategy he consists of two key phases. 1) Tool creation: LLMs (often called tool builders) create tools (implemented as Python functions) specifically for a particular job. 2) Tool Application: His second her LLM, known as Tool User, may be the same person as the creator of the tool, and applies the tool to address new requests. With a two-step design, LATM may assign work to the most appropriate LLM at each step.
In particular, powerful but resource-intensive models (such as GPT-4) have the potential to model competent processes of creating tools. On the other hand, lighter and more affordable models (such as the GPT-3.5 Turbo) are likely due to the significantly easier procedure of using the tool. This method significantly reduces the average computing cost of processing multiple jobs while improving the LLM’s problem-solving skills. For certain functions, he only has to go through the tooling steps once. Therefore, the created tool can be applied to several task instances.
This method offers a scalable and economical alternative for dealing with difficult problems. Consider a scenario where a user asks her LLM to arrange a meeting for everyone to attend (e.g. through an email exchange). Complex arithmetic reasoning problems are often difficult to complete on a lightweight machine like the GPT-3.5 Turbo. However, powerful models like GPT-4 are much more expensive to infer, but can get you the right answer. LATM overcomes these obstacles by taking a powerful but expensive model as tool creators and handing it over to a cost-effective model as tool users. After the tool has been forged, the user can utilize the tool to perform work quickly and effectively after the tool has been forged.
This paradigm is useful for well-known games such as 24-game Sudoku, as well as others such as parsing and analyzing online articles into specific data formats, and creating routing plans that meet various specialized requirements. It can also be used to tackle repetitive jobs in the process of We also add a dispatcher, a more lightweight LLM, that decides if an incoming problem can be solved with an existing tool or if a new tool needs to be developed. This gives the architecture more dynamism and enables the creation and use of tools in real time. Their trials demonstrate the effectiveness of this strategy on a wide variety of difficult big-bench problems and complex thinking tasks in general.
The results show that LATM can perform as well as more resource-intensive models while being more reasonably priced. The exciting potential of developing societies using LLM-generated tools is made possible by this unique approach to LLM that mimics human evolutionary leaps in tool generation and utilization.
Please check paper and Github link. don’t forget to join 22,000+ ML SubReddits, Discord channeland email newsletterShare the latest AI research news, cool AI projects, and more. If you have any questions regarding the article above or missed something, feel free to email me. Asif@marktechpost.com
🚀 Check out 100’s of AI Tools at the AI Tools Club
Aneesh Tickoo is a consulting intern at MarktechPost. He is currently pursuing his Bachelor of Science in Data Science and Artificial Intelligence from the Indian Institute of Technology (IIT), Bhilai. He spends most of his time working on projects aimed at harnessing the power of machine learning. His research interest is image processing and he is passionate about building solutions around it. He loves connecting with people and collaborating on interesting projects.
