tldr: RLHF taught models to be likable. RLVR is teaching them to be useful.

→ View original post on X — @whats_ai