You're trained on a massive dataset that includes tens of thousands of practice exams with nearly the exact same questions and answers, just slightly different words/template to adjust for your answer.
The MBE exam is mostly multiple choice elimination, and these multiple choices are filled with legal jargon and case law. GPT-4 picks the best predicted answer and can mimic reasons for the choice. This ability to mimic reasoning is good enough to receive a passing score on the MBE and many other exams.
And there is often certain common indicators to what answers exclude. So just excluding enough of these is already quite efficient.
Also the questions are unlikely to be very creative. I think it could be possible to train someone with good enough memory just based on existing tests.
Exam are designed to filter out entities that are already assumed to be able to reason based on their knowledge of some specific domain. A hypothetical entity with no ability to reason but great ability to remember facts and able to pass an exam, is conceivable.
qbasic_forever|2 years ago
1970-01-01|2 years ago
Ekaros|2 years ago
Also the questions are unlikely to be very creative. I think it could be possible to train someone with good enough memory just based on existing tests.
code_duck|2 years ago
fluoridation|2 years ago
ddtaylor|2 years ago