If exam questions are too easy, even poor students can score 100. In the AI community, how should we test the real abilities of the popular large models? With college entrance exam questions? Of course not!
Some believe that being first on various Benchmark lists means being the strongest. But that's not true. Sometimes, the more "authoritative" the list, the easier it is to game the rankings.
So, a model's strength isn't just about ranking first on a single Benchmark. It should perform well across multiple dimensions.
