rlhfopinionbearishRubrics are prone to over-optimization in a way similar to reward models, where RLVR (Reinforcement Learning from Verifiable Rubrics) is its own distinct phenomenonNathan Lambert26 Jul 2026https://x.com/natolambert/status/2081018219746468088