arXiv 2610.25060 · cs.LG · 2026-10-18
Policy Is Not Capability: Some Safety Benchmarks Measure Which Definition of "Unsafe" You Trained On
Vahid Jebraeeli, Jonathon Romero, Yijin Wang, Adrian Chan, Jianbang Zhang
机构待核实
一句话要点
安全基准实际上衡量的是训练时所采用的‘危险’定义,而非模型的真实能力,且其分数受标签定义显著影响,存在严重污染和混淆。
Safety benchmarks measure training policy definitions, not model capability, and are highly sensitive to label definitions.
要点
- 安全评估误将文本属性当作危险性,实则反映的是训练时所用政策定义,模型本身并无直接判断能力。
- 同一训练数据在不同标签标准下产生显著差异,十种安全模块在HarmBench和WildGuard上的得分差距达23倍,说明基准结果高度依赖标签定义。
- 仅0.03%参数变化即可导致基准分数大幅波动,表明模型表现受训练标签政策的微小扰动影响,存在严重可塑性问题。
- 验证准确率与标注者政策一致性强(相关系数0.74),而非检测能力,说明评估结果反映的是训练过程中的标签偏见,而非真实安全性能。
English overview
Safety benchmarks do not evaluate model capability but reflect the training policy's definition of 'unsafe'. The study shows that benchmark scores vary dramatically based on label definitions, with significant contamination from training data. Performance differences across models are driven by policy labels rather than detection ability, and validation aligns more with annotator policy than actual safety detection.
只根据 arXiv 官方摘要,由开源模型整理,经数字/专名自动核对(必须在摘要中出现),以论文原文为准。
阅读 arXiv 原文 PDF Hugging Face 讨论
主题与学科
同主题:安全与可解释
- Behavioral Compositionality of Sparse Autoencoder Representations ▲108
- DecepEval: A Benchmark for Evaluating Deception in LLM Agents ▲85
- U-Space: Uncovering When and Why Uncertainty Arises in Language Models ▲54
- Lexical Reference Frames: Interpreting LM Residual Streams in Native Semantics ▲46
- Just Ask Jev: Reinforcement Learning for Calibrated Decisions as a Zero-Shot Detector of AI Alignment Failures ▲29
同主题:评测与数据
- Spectral-Spatial Interpretation 2026-10-15 · ▲753
- StreamEval: Evaluating Interactive Streaming Video Generation Models 2026-10-20 · ▲649
- Has GPT-6-Astra Achieved Embodied Intelligence? 2026-10-15 · ▲421
- BizAgentBench: Unveiling Hidden Challenges in Deploying Long-Horizon Agents for Real-World Business 2026-10-14 · ▲388
- RealCompanion: Benchmarking Human Understanding from Reasoning over Longitudinal Real-World Conversations 2026-10-01 · ▲287
数据说明
元数据来自 arXiv API(已核实);点赞来自 Hugging Face 每日论文;引用来自 Semantic Scholar。最近刷新 2026-10-18。
要点由本站根据摘要自写,不转载摘要原文,请阅读原文。