Tool Up! DeepMind, Princeton & Stanford’s LATM Enables LLMs to Make Their Own Tools

The 19th-century British philosopher Thomas Carlyle ascribed human progress to a key historical development: “Man is a tool-using animal. Without tools he is nothing, with tools he is all.” While today’s large language models (LLMs) have demonstrated impressive generative and problem-solving capabilities, recent research suggests they could take a similar leap forward through the use of external tools. Research in this area has thus far focused on existing tools such as calculators and search engines — but could LLMs learn to address emerging challenges by fabricating their own tools?

In the new paper Large Language Models as Tool Makers, a research team from Google DeepMind, Princeton University and Stanford University presents LATM (large language models as tool makers), a closed-loop framework that enables LLMs to create their own reusable tools to boost efficiency and enhance their problem-solving capabilities.

The LATM approach comprises two stages: 1) Tool making, where an LLM crafts tools for given tasks via a Python utility function, and 2) Tool Using, where the same or another LLM applies these tools for solving new requests. This setup enables LATM to allocate jobs to the most suitable LLM at each stage — for example, a powerful but computation-intensive LLM might be called on in the complex tool-making stage to ensure high accuracy, while a lightweight and more cost-effective model could fill the role in the simpler tool-using stage. As such, LATM can not only improve LLMs’ problem-solving capabilities, but also substantially cut total computational costs.

Notably, the generated tools can be reused for recurring tasks in a workflow or even applied to new tasks as appropriate, enabling the development of scalable and cost-efficient LLMs for addressing complex tasks.

The team also introduces the “dispatcher,” an additional lightweight LLM designed to determine whether a given problem can be solved using currently available tools or if a new tool should be created.

In their empirical study, the team applied LATM on complex reasoning problems such as Big-Bench tasks. In the experiments, LATM achieved performance comparable with resource-intensive models such as GPT-4 while significantly cutting computation costs.

This work opens a promising new path for improving LLMs’ advanced problem-solving skills by enabling them to create their own tools for a given job. The team suggests an exciting future research direction might involve enabling their tool maker to refine and upgrade existing tools to more efficiently address new tasks.

The code is available on the project’s GitHub. The paper Large Language Models as Tool Makers on arXiv.

---

Author: Hecate He | Editor: Michael Sarazen

---

We know you don’t want to miss any news or research breakthroughs. Subscribe to our popular newsletter Synced Global AI Weekly to get weekly AI updates.