A smarter way to test AI on feedback
A new benchmark tests whether AI models actually improve when given feedback after wrong answers, not just whether they get things right first.
A new benchmark tests whether AI models actually improve when given feedback after wrong answers, not just whether they get things right first.