1
LLM Judges Inconsistently Disagree Across Safety Criteria and Harm Categories
LLM作为安全评估者存在系统性不一致：跨越不同安全标准和危害类别时的分歧模式揭示了AI对齐评估的脆弱性。
arXiv:2605.31381v1 Announce Type: new Abstract: We evaluate the consistency of automated judges in conducting a multi-dimensional safety evaluation in…