This is especially true for public benchmarks, which model makers can easily hill climb by simply creating a bunch of RL environments that closely mimic the benchmark tasks.
作者给自己结论留的后门，也是最该记住的一条：公开基准可以用定制 RL 环境直接刷分。于是"开源已追上"的证据强度取决于评测是否私有，而本文所用基准恰恰多为公开——结论方向可信，幅度存疑。