AI Model Evaluation
Swap the question set and the same model moves several places — a leaderboard is never a report card, it is one slice taken from the examiner's angle. How benchmarks are built, how graders are designed and how scores stop being inflated decide whether a number can be trusted.
Evaluation's hard part is not writing questions but making scores reflect ability. UNO-Bench unifies 44 task types with a scoring model at about 95 percent agreement; AMO-Bench has experts write 50 fresh IMO-level problems to dodge memorization, and most models still score under 40 percent. Agent evaluation starts from real failures and separates pass@k from pass^k — only the latter means reliable every time.
