Bridging Knowledge Gaps in LLMs via Function Calls

Abstract

Large Language Models (LLMs) demonstrate impressive abilities across a wide range of NLP tasks. However, their underlying architecture and design come with inherent limitations, which result in issues like hallucinations and constrained reasoning capabilities. Additionally, creating an autonomous AI agent capable of handling complex real-world tasks demands access to real-time information, sensitive data, or external tools-capabilities that most LLMs currently lack. Addressing these issues may require augmenting LLMs with external knowledge through function calling. These function calls serve as an interface between LLMs and the world, enabling access to real-time data, diverse tools, reasoning systems, knowledge graphs, APIs, plugins, code interpreters, and more. The primary objective of this talk is to highlight the significance of function-calling capabilities in bridging the knowledge gap in LLMs, showcase recent research advancements in this area, and discuss existing challenges along with future directions. Also, I will present a training and benchmarking data suite for function calling - API-BLEND and a function calling model - Granite-20B-FunctionCalling.