arXiv
Language Models
Why AI stumbles on tricky probability — even when it aces the basics
LLMs are like chess grandmasters who play flawlessly in normal games but choke the moment you change the board layout slightly — they've memorized patterns, not principles.
This means the models we're using for reasoning and decision support have a fragile grasp of probability that collapses under small prompt rewording, so we can't trust them blindly for anything involving chance or counterintuitive logic.
Bug reported: No