Researchers have introduced Skill-RM, a novel approach to unifying heterogeneous evaluation criteria for large language models (LLMs) via agent skill. This development addresses the current reliance on disparate criteria, such as rule-based verifiers and complex rubrics, which can hinder the effectiveness of reinforced fine-tuning and reinforcement learning pipelines1. By integrating various types of evidence, Skill-RM provides a more comprehensive and unified mechanism for evaluating LLMs. This is particularly significant in the context of LLM development, where reinforcement learning can reshape both capability and risk surfaces. As LLMs become increasingly powerful, their security implications cannot be ignored, and a unified evaluation framework is essential for mitigating potential risks. The introduction of Skill-RM has important implications for practitioners, as it can help to streamline the evaluation process and improve the overall security and reliability of LLMs.