Google

The process of tool use by an LLM involves generating a specially formatted string, such as `{tool: calculator, expression: 10 * 4 * 2}`, which acts as a trigger for an external tool like a mathematical interpreter. Once the tool computes the result (e.g., '80'), this output replaces the original tool-invocation string. This substitution updates the model's working context, enabling the result to be directly referenced and utilized in the subsequent token predictions that form the next steps of the reasoning process.

Mechanism of Tool Invocation and Result Substitution

An AI assistant is solving the problem: 'A company has 4 departments, and each department needs 10 new computers. If each computer costs $950, what is the total cost?' The assistant's internal reasoning process is captured at two sequential points in time. Analyze the change between 'Context Before' and 'Context After' and explain the specific mechanism responsible for this transformation.

Analysis of LLM Context Modification

A large language model needs to answer the question, 'What is 5 times 12?'. To do this, it uses an external calculation tool. Arrange the following steps in the correct chronological order to show how the model generates the final answer.

An AI assistant is tasked with solving a multi-step problem: 'Calculate the area of a rectangle with a width of 10 and a length of 4, and then multiply the result by 2.' The assistant's internal process is as follows:

1. It first generates the string: `{tool: calculator, expression: '10 * 4'}`.
2. An external system intercepts this string, computes the expression, and obtains the result '40'.
3. The system then updates the assistant's working context before the assistant generates its next thou

Learn Before

Related