A new framework for editing what LLMs 'know' without breaking their reasoning

Researchers propose principled methods for editing facts inside LLMs without degrading their reasoning ability.