[论文] What Does a Bias Head Measure? Non-Identifiability in Head-Level Bias ...
论文概要
研究领域: NLP
作者: S. M Seefat Alam
发布时间: 2026-11-17
arXiv: 2611.23218
中文摘要
将 LLM 社会偏见定位到单个注意力头并通过消融移除的方法,仅基于公平性指标的影响验证,从未证明选择这些头的统计量确实测量了偏见。我们在 Qwen-2.5-7B 和 Llama-3.1-8B 上审计了该方法。头图与无视标签的响应长度图相关性达 ρ=-0.950 和 -0.768,表面识别出混淆因素。打破长度-标签耦合的干预拒绝了这一解读:提示固定时,响应长度仅解释方差的 1%,而提示身份解释 64%。在任何固定提示内,公平标签、响应长度和受保护群体标记的存在在 φ=0.95 处共线(部分由标记规则定义决定),任何观察性分析都无法归因。实质性声明独立失败:偏见选择性不集中于晚期层,所有测试头集消融后均未降低不公平率;但其补集在相反标签评分下可靠地破坏拒绝行为(p=2.7×10⁻⁵),表明该流程在此尺度上能检测到效应。
原文摘要
Methods that localise social bias in large language models to individual attention heads and remove it by ablation are validated by their effect on a fairness metric, never by evidence that the statistic selecting the heads measures bias. We audit this on Qwen-2.5-7B-Instruct and Llama-3.1-8B-Instruct. The head map correlates at \(ρ= -0.950\) and \(-0.768\) with a label-blind response-length map, apparently identifying a surface confound. An intervention that breaks the length-label coupling rejects this reading: with the prompt held fixed, response length explains 1% of the statistic's variance. What does explain it is prompt identity, at 64% of variance, in an ordering that is not monotonic in length. Within any fixed prompt the fairness label, response length and the presence of a protected...
自动采集于 2026-11-19
#论文 #arXiv #NLP #小凯