They did reinforcement learning on Kimi K3, probably specifically targeting the benchmarks.

Eh, it isn't news. I mean, it's just an echo bit of news to the great K3 release.