跳到正文
研踪PaperTrail
← 研踪 PaperTrail

arXiv:2611.23218 · v1

What Does a Bias Head Measure? Non-Identifiability in Head-Level Bias Attribution

偏见头测量的是什么?头级偏见归因中的不可识别性

首次提交 2026-11-18 02:57
版本更新 2026-11-18 02:57
雷达观测 2026-11-18 12:34
arXiv ↗PDF ↗

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=−0.950ρ= -0.950 and −0.768-0.768 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-group token are collinear at φ=0.95φ= 0.95, partly by definition of the labelling rule, so no observational analysis can attribute the map to any of them, and no method in this family performs an experiment that could. The substantive claims fail independently: bias selectivity is not late-layer-concentrated on either model under either reference reading, and no head set we tested reduces the unfairness rate when ablated. Its complement, scored on the opposite label, reliably destroys refusal behaviour (p=2.7×10−5p = 2.7\times10^{-5}), demonstrating that the pipeline detects effects at this scale.

作者备注:18 pages, 13 tables. Under review at ACL Rolling Review (October 2026 cycle)

AI 论文解读

基于摘要 · 非全文审阅

作者审计头级偏见归因方法,发现其统计量主要由提示身份驱动,且无法区分标签、长度与受保护群体标记,同时偏见选择性并非晚层集中,消融头集也未降低不公平率。

正在读取账号与研究记录…

版本差异

目前只收录了一个版本。收录下一版后,可以对照标题、作者、分类和摘要。

观测到的版本

相邻记录可能跳过中间版本;首次发现 v2 及以上也会标为版本更新。

v1 · 2026-11-18 02:57 · 最新观测版本

What Does a Bias Head Measure? Non-Identifiability in Head-Level Bias Attribution

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=−0.950ρ= -0.950 and −0.768-0.768 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-group token are collinear at φ=0.95φ= 0.95, partly by definition of the labelling rule, so no observational analysis can attribute the map to any of them, and no method in this family performs an experiment that could. The substantive claims fail independently: bias selectivity is not late-layer-concentrated on either model under either reference reading, and no head set we tested reduces the unfairness rate when ablated. Its complement, scored on the opposite label, reliably destroys refusal behaviour (p=2.7×10−5p = 2.7\times10^{-5}), demonstrating that the pipeline detects effects at this scale.

查看此版本解读 →查看此版本原文 ↗