arXiv cs.AI· Joseph Egan, Ciar\'{a}n M. Gilligan-Lee, Michael O'Riordan, Yuchen Zhu·· 6 天前AI 评分32
无需监督即可发现潜意识学习：蒸馏模型隐性特质的防御新方法
Defence against the dark arts of distillation: Unsupervised discovery of subliminal learning
AI 导读
研究者提出一种无需监督的方法，仅凭参数更新即可从蒸馏模型中恢复被潜意识传递的行为特质，利用特质在教师模型生成的多样数据集上表现一致、而任务能力各异的非对称性。该方法在 Cloud et al. (2026) 的潜意识学习基准上，无需任何特质标签就恢复了传递的动物偏好特质，并能抑制其表达且几乎不影响任务性能。
正文 · 原文
Abstract:When a language model is distilled from a teacher that shares its base parameters, the student can inherit behavioural traits of the teacher---including misalignment---even from data that is semantically unrelated to the trait (Cloud et al., 2026). All existing defences require knowing the trait. We remove this requirement and prove that a subliminally transferred trait can be recovered from the parameter updates alone, without supervision, by exploiting a single asymmetry: the trait is expressed consistently across diverse datasets generated by the teacher, whereas task-specific and shared capabilities vary. The resulting estimator returns a weight-space direction that can be amplified to discover and interpret the trait or projected out to suppress it. Applied to the subliminal-learning benchmarks of (Cloud et al., 2026), our method recovers the transmitted animal-preference traits without any trait labels and suppresses their expression with almost no impact to task performance.
| Comments: | 9+12 pages, comments welcome |
| Subjects: | Machine Learning (stat.ML); Artificial Intelligence (cs.AI); Machine Learning (cs.LG) |
| Cite as: | arXiv:2610.13536 [stat.ML] |
| (or arXiv:2610.13536v1 [stat.ML] for this version) | |
| https://doi.org/10.48550/arXiv.2610.13536 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Ciarán M. Gilligan-Lee [view email]
[v1]
Fri, 9 Oct 2026 12:55:50 UTC (222 KB)
来源：arXiv cs.AI · arxiv.org