TamperBench

档案区分来源方法阅读与既有目录信息；已经公开不代表完成独立证据复核或双语编辑复核。

测量构念

开放权重防护对保留效用的修改有多稳健？

任务与样本

21个模型面对九类篡改，并分别搜索超参数，同时评价安全与效用。

判分与失效情形

MMLU-Pro相对退化≤10%时最坏StrongREJECT，攻击均值另计。

系统条件

主要搜索每模型攻击对40次超参数试验。

统计含义

来源报告的结果表述；这不等同于已确定运行次数、分母或估计方法： 主要比较在MMLU-Pro相对退化不超过10%时选择最坏StrongREJECT结果。 每模型攻击40次搜索定义了对抗预算。

效度与解释

主要模型为0.6–8B，聚焦拒绝型防护，大模型覆盖初步（§5）。

适用研究问题

模型建议：比较时保留效用约束与攻击搜索预算。

• TamperBench : Systematically Stress-Testing LLM Safety Under Fine-Tuning and Tampering · AbstractSelected original resource https://arxiv.org/html/2602.06911; sha256:791d59fea870ef2288ea6f4c60dd83b68233441d4eeafef2b3aabc449a40b22a; archived source capture; exact capture time not independently established
• TamperBench : Systematically Stress-Testing LLM Safety Under Fine-Tuning and Tampering · §4 Experiments: scoring and 40-trial searchSelected original resource https://arxiv.org/html/2602.06911; sha256:791d59fea870ef2288ea6f4c60dd83b68233441d4eeafef2b3aabc449a40b22a; archived source capture; exact capture time not independently established
• TamperBench : Systematically Stress-Testing LLM Safety Under Fine-Tuning and Tampering · §5 LimitationsSelected original resource https://arxiv.org/html/2602.06911; sha256:791d59fea870ef2288ea6f4c60dd83b68233441d4eeafef2b3aabc449a40b22a; archived source capture; exact capture time not independently established