Peeking Inside LLMs Could Boost AI Safety, Researchers Say

Researchers propose identifying cognitive elements in large language models that signal potential unwanted actions. This approach aims to improve AI safety by understanding internal model behavior rather than just outputs. The method could help predict and prevent harmful AI decisions before they occur.