Concurrency without Model Changes: Future-based Asynchronous Function Calling for LLMs

要約

Function calling, also known as tool use, is a core capability of modern LLM agents but is typically constrained by synchronous execution semantics. Under these semantics, LLM decoding is blocked until each function call completes, resulting in increasing end-to-end latency. In this work, we introdu…