The paper finds that even sophisticated large language models are frequently stumped by a simple logic question termed the 'Alice in Wonderland' problem, casting doubt on the claimed advancement of AI models.
Despite the seemingly straightforward question requiring basic reasoning, top AI language models like GPT-3, GPT-4, and others struggled to provide accurate answers, showcasing a significant gap in logic processing abilities.
Impressively, only the new GPT-4o model achieved a passing success rate on this logic question, while other models not only failed to answer correctly but also displayed bizarre and incorrect lines of reasoning when explaining their answers.
Collection
[
|
...
]