Why LLMs are vulnerable to the ‘butterfly effect’

Researchers found that the smallest of perturbations, such as adding a space at the end of a prompt, can cause an LLM to change its answer.

Posted from: this blog via Microsoft Flow.