Reproduce “Beyond Binary Labels: Diagnostic Evaluation of LLM Safety with Prompt-Level Rubrics” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Hongyu Li, Xinlong Huang, Zhan Yu, Jing Zhao, Zhe Zhao, Juntao Zhong, Xiong Kuang, Rui Yuan and 8 more
DiagSafe evaluates language-model safety with prompt-specific rubrics that separate harmful behaviors, exemptions, and usefulness. Its bilingual benchmark contains 3,489 prompts, including instruction attacks that embed harmful intent in ordinary tasks. On 236 human-annotated responses, Hy3’s rubric-guided agreement was 95.34%, versus 87.29% with a generic prompt; all 11 locally served open-weight models had higher unsafe rates on DiagAttack than DiagBase.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.