DiagSafe: prompt-level rubrics for diagnosing LLM safety

1. 17:48
   XiaohongshuXiaohongshu · AI Security Heart· Discussion

   LLM safety diagnostic evaluation based on prompt-level scoring criteria(opens the original post in a new tab)

1. 13:33
   arXiv: Alignment, Deception & Oversight

   Beyond Binary Labels: Prompt-Level Rubric-Based Diagnostic Evaluation of LLM Safety

   The researchers propose DiagSafe, a diagnostic evaluation framework that replaces binary safe/unsafe labels with prompt-level rubrics to evaluate LLM safety; the rubrics are based on globally applicable AI safety policies and give explicit criteria along two dimensions: safety (targeting specific harms and setting exemptions to suppress false positives) and helpfulness (independently assessing instruction completion). The accompanying benchmark is bilingual in Chinese and English, based on a taxonomy of 84 safety categories, and contains two subsets: DiagBase covers single-turn and multi-turn directly harmful queries and templated jailbreaks, and DiagAttack embeds harmful intent into complex instruction-following tasks to form instruction attacks. Experiments show that the framework achieves 95.34% agreement between LLM-as-a-judge and human annotation, higher than the 69%-87% of baseline methods, with an inter-judge Cohen's κ of 0.864. Applying the framework also finds that mainstream LLMs differ markedly in unsafe rates, that each model has unique weaknesses across different dimensions, and that they are especially susceptible in Chinese and multi-turn scenarios.

Current attention 33·Peak in the comparable period 33 (Nov 6 03:00)·Change over the comparable past 24 hours 0%