The Evolution of Large Language Models
Large language models (LLMs) like GPT-3, Codex, and GPT4 have been pushing the boundaries of artificial intelligence. But they’re not without their flaws. They struggle with recognizing current information, providing precise mathematical solutions, and reasoning across complex logic chains. Researchers have been working on external tools to enhance their abilities, but these tools have their own challenges.
Introducing CREATOR
Researchers from Tsinghua University and the University of Illinois have introduced CREATOR, a framework that empowers LLMs to be the developers of tools. Unlike traditional tool-using frameworks, CREATOR focuses on diversifying the toolset and improving resilience and correctness.
Four Steps of CREATOR
- Creation: LLMs use abstract reasoning to create broadly applicable tools.
- Decision: LLMs decide when and how to apply the tools.
- Implementation: LLMs run the program to solve the problem.
- Rectification: LLMs alter the tools and choices based on the outcomes.
Performance and Impact
CREATOR has been tested on existing benchmarks like MATH and TabMWP, outperforming traditional methods. A new dataset called the Creation Challenge has also been introduced to evaluate tool creation. The findings show that tool development encourages knowledge transfer and enables LLMs to adapt more effectively to various contexts.
