A new benchmark called Preference Proxy Evaluations (PPE) has been introduced to evaluate reward models (RMs) and large language models (LLMs) in the context of reinforcement learning from human feedback (RLHF). The benchmark addresses critical questions regarding the effectiveness of RMs in guiding RLHF and whether LLM judges can replace human evaluations. PPE utilizes real-world human preferences gathered from Chatbot Arena, encompassing over 16,000 prompts and 32,000 diverse model responses. The initiative, led by Evan Frick and his team, aims to provide a framework for meta-evaluation of these models, thereby facilitating better model alignment and performance assessment in AI systems. This development follows ongoing discussions within the AI community about improving evaluation methodologies for machine learning models.