It’s like LLMs freeze up at the slightest exception, like they’ve never seen anything go off-script before. Seriously though, exceptions are part of the game, they’re how we learn and improve. We need to give these models some better coping mechanisms for those edge cases