Mode 7: LLM for Tool Usage
Definition
To enhance LLMs with additional capabilities when they lack confidence in answering questions—such as web searching, running code, or even processing multimodal data—external system components can be integrated. These components extend the LLM's functionality, allowing it to interact dynamically with various tools and data formats for more comprehensive responses. These interactions often occur during the generation stage, where the LLM not only produces user responses but also generates function call commands, leveraging specific arguments provided by the model itself. The system proposed by Lehmann et al. establishes a tool set enabling LLMs to combine data from diverse sources as a structured library.
Figure: How external tools are integrated during the Generating and Evaluating stages.
Key Interaction Modes and Examples
User uses prompting techniques to guide LLM in external tools usage.
- Explicit reasoning. To improve the effectiveness of external tool usage in question answering, Zhang et al. proposed a hierarchical, step-by-step prompting method that decomposes a claim into multiple subclaims, each verified through sequential question-answering steps.