arXiv cs.CL· Chaemin Jang, Donggyu Lee, Jeesoo Kim, Dongman Lee, Jihee Kim·· 2 天前AI 评分32
用无标签子空间分离语言模型的判断与表达
Isolating a Language Model's Judgment with a Label-Free Subspace
AI 导读
研究者发现，同一输入在不同指令下隐藏状态向量的差异集中在一个低维"表达子空间"中，无需基准标签即可识别该子空间，并从线性预测器中移除其贡献以获得"孤立判断"。在正确性、蕴含和安全任务上，孤立判断在未见指令下仍保持可靠；当判断问题被否定时，蕴含任务 AUROC 从 0.504 升至 0.825，安全任务从 0.480 升至 0.891。判断在表达子空间干扰预测器之前的更早层即可读出。
正文 · 原文
Abstract:Language models are often asked to make judgments: whether a proposed answer is correct, whether one statement follows logically from another, or whether a response is safe. Yet changing the instructions that specify how the judgment should be expressed can make the final answer less reliable, even for the same input. We ask whether this reflects a change in internal judgment or a failure to express it. Comparing hidden-state vectors for the same input under different instructions, we find that their differences concentrate in a low-dimensional expression subspace. We identify this subspace without benchmark labels and remove its contribution from a linear predictor to obtain an isolated judgment. Across correctness, entailment, and safety, the isolated judgment remains reliable under instructions not used to identify the subspace, even when the model's final answer and the original predictor degrade. When the judgment question is negated, isolation raises the predictor's AUROC from 0.504 to 0.825 for entailment and from 0.480 to 0.891 for safety. We also find that judgment is readable in earlier layers, before the expression subspace begins to disrupt the predictor. Together, these results provide a way to separate judgment from expression and recover internal judgments even when the model's final answers are unreliable.
| Subjects: | Machine Learning (cs.LG); Computation and Language (cs.CL) |
| Cite as: | arXiv:2610.21256 [cs.LG] |
| (or arXiv:2610.21256v1 [cs.LG] for this version) | |
| https://doi.org/10.48550/arXiv.2610.21256 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Chaemin Jang [view email]
[v1]
Thu, 15 Oct 2026 04:49:49 UTC (118 KB)
来源：arXiv cs.CL · arxiv.org