1 article.
MMLU and similar multiple-choice benchmarks no longer discriminate between strong models. Task-based, real-world evaluations like SWE-bench give a more honest picture of capability.