HARP: Measuring Harm Amplification in Multi-Agent LLM Systems

A new methodology called HARP (Harm Amplification through Role Perturbation) has been introduced to study harm amplification in multi-agent LLM systems, which decompose workflows across various components. HARP aims to measure local and global harm by comparing clean and perturbed executions, focusing on how perturbations can propagate and amplify risks within the system.

This development is significant as it enhances the interpretability of multi-agent systems while addressing the potential risks of harm amplification, which is crucial for ensuring the safety and reliability of AI applications in complex environments.