The researchers argue that traditional benchmarks, like math and coding tests, are flawed due to “data contamination” and ...