The article presents a visualization approach to explore the statuses and behaviors of agents within Large Language Model-based Autonomous Systems (LLMAS). A pipeline is proposed that forms a behavior structure from raw LLMAS execution events, applies a behavior summarization algorithm to create a hierarchical summary of the entire structure in time sequence, and uses a cause trace method to discover the causal relationship between agent behaviors. AgentLens, a visual analysis system, is developed that uses a hierarchical temporal visualization to illustrate the evolution of LLMAS and allows users to interactively investigate the details and causes of agents’ behaviors.
Publication date: 15 Feb 2024
Project Page: ?
Paper: https://arxiv.org/pdf/2402.08995
