Large scale language models (LLMs) have made significant progress in recent years, including GPT-3, Codex, PaLM, LLaMA, ChatGPT, and the newer GPT4. The potential of LLM is getting closer and closer to artificial general intelligence thanks to the excellent performance of these models on in-context learning, code generation, and various other NLP tasks. Despite these impressive achievements, current LLMs suffer from the inability to perceive or react to current information, their frequent inability to provide precise and comprehensible mathematical solutions, and their inability to span long logic chains. It has drawbacks such as unstable inference. A body of research is motivated to provide LLMs with external tools to reduce the burden of memorization and improve their ability to solve problems. For example, by incorporating tools such as web search engines and question-and-answer (QA) systems, LLMs can learn when and how to use external resources for problem solving. Recent research also uses additional external LLM tools such as GitHub resources, neural network models (such as the Huggingface module), and code interpreters (such as the Python interpreter). LLMs must provide a broad blueprint before using these techniques to solve complex problems.
Nevertheless, tool-enhanced LLM still faces some difficulties and pays particular attention to the following areas: (1) the diversity of potential innovative tasks is intrinsic; remains endless, but most of our current work is focused on a few tools. As a result, it can be difficult to find existing tools that are suitable for solving new problems. (2) Current approaches in language models to estimate how to use tools most effectively are inherently complex. The entire task-handling process requires extensive planning, which imposes a large cognitive burden on the model and requires high training costs. (3) After receiving execution results, the tooling pipeline has no defined automated error handling mechanism. Further development is needed on the accuracy and robustness of the framework. In this study, researchers from Tsinghua University and the University of Illinois (UC) aim to approach these disorders from a new perspective. LLMs become the developers of tools that enable them to solve problems with greater precision and flexibility. Rather than having the LLM act as a tool consumer.
As a result, the company introduced CREATOR, a tool development framework. It uses the power of LLM to develop tools and modify them according to existing parameters before addressing a particular problem. Figure 1 shows the pipeline variations between CREATOR and common tooling frameworks. The Tool Usage Framework focuses on how to use reasoning to more effectively select and plan API usage. In contrast, they diversify their toolset, separate different levels of rationale, and focus on improving the framework’s resilience and correctness.
CREATOR is divided into four steps.
• Create: LLM’s ability to make abstract reasoning based on problems is used to create broadly applicable tools through documentation and code realization.
• Decision: Use the appropriate tool and choose when and how to apply the tool.
• Implementation: The LLM runs a program that uses the tools to address the problem.
• Modification: Change the means and options based on the results of the run.
First, run tests in CREATOR using MATH and TabMWP as two existing benchmarks to see how successful the design is. TabMWP offers a variety of tabular settings for problem solving, while the MATH dataset contains difficult and diverse math competition challenges. “Notably, ChatGPT built on CREATOR significantly outperforms traditional Chain of Thought (CoT), Program of Thought (PoT), and tool usage baselines, achieving average accuracies of 59.7% and 94.7%, respectively. For the MATH and TabMWP datasets.
They also propose the Creation Challenge dataset. It consists of innovative and difficult questions that need to be answered with existing tools or code packages, as existing benchmarks were not specifically designed to evaluate tool creation. We will use this dataset to demonstrate the value and use of LLM’s tool-building capabilities. Additionally, how tool development facilitates the transfer of knowledge, experimental results and case studies show that LLMs have varying degrees of tooling proficiency, which allows them to adapt more effectively to different problem situations. is also provided.
please check out paper. don’t forget to join 23,000+ ML SubReddit, 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 us. 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 in image processing and he is passionate about building solutions around it. He loves connecting with people and collaborating on interesting projects.
