They're clearly building better training datasets and doing extensive RL on these benchmarks over time. The out of distribution performance is still awful.