Researchers have introduced a new benchmark called VARM-Bench to evaluate the performance of language models in moderating Chinese online content. The benchmark focuses on providing verifiable and structured reasoning for moderation decisions, which is crucial for ensuring the reliability and transparency of online content moderation. Unlike existing benchmarks that focus on label classification or fine-grained toxicity categorization, VARM-Bench requires models to provide a clear and explicit rationale for their moderation decisions. This includes identifying targets, determining target types, and assessing author stances. The benchmark evaluates language models under various protocols, including zero-shot prompting and structured CoT supervision, and highlights the importance of auditing and reproducibility in online content moderation. --- Why it matters: This matters to researchers in AI because it provides a new standard for evaluating the performance of language models in critical applications like online content moderation. By requiring models to provide transparent and verifiable reasoning, VARM-Bench can help improve the reliability and accountability of these systems. Source: https://arxiv.org/abs/2608.15600