Some text is shown in its original language while the English translation is being prepared.
BackSource arXiv
Leo Schwinn, Moritz Ladenburger, Tim Beyer, Mehrnaz Mofakhami, Gauthier Gidel, Stephan Günnemann
Indexed by this site 2026-10-12 14:28 2 hr ago
Originally published Feb 4
Selected 2026-10-12 14:39
Category Evaluation arXiv · arXiv:2603.06594 · Leo Schwinn, Moritz Ladenburger, Tim Beyer, Mehrnaz Mofakhami, Gauthier Gidel, Stephan Günnemann · Indexed by this site 2026-10-12 14:28 · Originally publishedFeb 4 SelectedAttention61 安全评测如同抛硬币:LLM 评委无法可靠衡量对抗鲁棒性 A Coin Flip for Safety: LLM Judges Fail to Reliably Measure Adversarial Robustness
AI reading guide
慕尼黑工业大学、Mila 等机构的研究者用 6642 条人工标注审计 LLM-as-a-Judge,发现其在对抗鲁棒性评测中平均仅略好于随机猜测。研究指出攻击、受害模型和语义类别三类分布偏移会叠加,使评委准确率降到接近随机水平,JailJudge 在 Llama-3.1-8B 上的 AUROC 最低为 0.48。作者称 BoN 等攻击的高成功率更多来自评委假阳性而非真正有害输出,按评委精确率修正后攻击成功率明显下降。为此他们发布 ReliableBench 和 JudgeStressTest 两个数据集,前者选取 41 个最易评判的行为可将平均准确率提升到 70%,后者包含 971 条评委普遍失败的样本。
Deep reading Generating
Why it was selected
论文以 6642 条人工标注审计 LLM 评委,显示其在对抗评测中接近随机猜测,并给出按评委精确率修正攻击成功率的做法。
Read original arxiv.org Why it was selected SelectedAttention61
Attention 0–100: content value contributes up to 70 and circulation up to 30; circulation decays over time. How it was rated
论文以 6642 条人工标注审计 LLM 评委,显示其在对抗评测中接近随机猜测,并给出按评委精确率修正攻击成功率的做法。