rlhfopinionbearishRubrics used in RLVR are prone to over-optimization in a way similar to reward modelsNathan Lambert26 Jul 2026https://x.com/natolambert/status/2081018219746468088