ResearchMarch 24, 2025•Updated March 26, 2025•18 min read
Emergent Tool Use in GPT-4
Systematic analysis of how language models spontaneously learn to use external tools with minimal prompting.
How do language models learn to use tools? In this article, I analyze the surprising ways GPT-4 adapts to new tasks with minimal instruction.
The Experiment
I gave GPT-4 access to a calculator, web search, and code execution tools. With just a few examples, it learned to use them effectively.
Prompt Example
User: What is 17 * 23?
Assistant: [Calculator] 391
Key Findings
- GPT-4 generalizes tool use from very few demonstrations
- It can chain tools together for complex tasks
- Failures often reveal gaps in world knowledge, not tool logic
"The most impressive part was how quickly GPT-4 adapted to new tools, even ones it had never seen before."
Priya Nair
AI Researcher