NUWA 风险发现 · 研究
Subliminal Learning: Language models transmit behavioral traits via hidden signals in data
NUWA 研究解读 · 独立复核待完成
NUWA Research Note
研究问题研究设计与假设来源发现NUWA 解读局限与相反证据研究联系研究与决策意义机构与角色原始资源模型辅助整理 · 选读原始全文或官方文章中的相关章节；尚未完成逐章节审计。
研究速读
即使过滤后的输出没有明显语义线索，蒸馏仍可能传递行为特征。
研究问题
教师模型的行为特征能否经语义无关的训练输出传给学生？
研究设计与假设
教师经提示或微调获得特征，共享参考模型的学生从过滤数字、代码或推理学习；跨家族对照与理想化定理检查机制。
来源发现
多个同基座设置发生特征传递。
所测跨家族传递未成功，语义分类器也不能可靠识别过滤数据中的被传特征。
NUWA 解读
NUWA 解读：仅做合成数据语义审查不足以确立蒸馏行为中立。
局限与相反证据
任务与提示较人工；定理假设共同初始化和小更新，不适用于任意教师、学生或特征。
研究联系
NUWA 关联：Beyond Owls 测试超越简单偏好的能力与条件行为；智能体自修改加入训练数据选择的部署后果。
研究与决策意义
NUWA 建议：保留师生谱系，内容过滤通过后仍检查蒸馏引发的行为变化。
精确来源位置
- Subliminal Learning: Language models transmit behavioral traits via hidden signals in data · 2 Experimental setup: distillation on an unrelated domainSelected original resource https://arxiv.org/html/2507.14805; sha256:2170110f6235ea5c9d6e80c605a198863a337db14ff9d0cf926644fa4a66db8c; archived source capture; exact capture time not independently established
- Subliminal Learning: Language models transmit behavioral traits via hidden signals in data · 3 Models transmit traits via numbersSelected original resource https://arxiv.org/html/2507.14805; sha256:2170110f6235ea5c9d6e80c605a198863a337db14ff9d0cf926644fa4a66db8c; archived source capture; exact capture time not independently established
- Subliminal Learning: Language models transmit behavioral traits via hidden signals in data · 5.1 Cross-model transmissionSelected original resource https://arxiv.org/html/2507.14805; sha256:2170110f6235ea5c9d6e80c605a198863a337db14ff9d0cf926644fa4a66db8c; archived source capture; exact capture time not independently established
- Subliminal Learning: Language models transmit behavioral traits via hidden signals in data · 6 Subliminal learning as a general phenomenonSelected original resource https://arxiv.org/html/2507.14805; sha256:2170110f6235ea5c9d6e80c605a198863a337db14ff9d0cf926644fa4a66db8c; archived source capture; exact capture time not independently established
- Subliminal Learning: Language models transmit behavioral traits via hidden signals in data · 8 DiscussionSelected original resource https://arxiv.org/html/2507.14805; sha256:2170110f6235ea5c9d6e80c605a198863a337db14ff9d0cf926644fa4a66db8c; archived source capture; exact capture time not independently established
来源核查、数据提取、编辑批准与实验复现分别记录。 方法说明 →
机构与角色
作者 (9)
Alex Cloud · Minh Le · James Chua · Jan Betley · Anna Sztyber-Betley · Sören Mindermann · Jacob Hilton · Samuel Marks · Owain Evans
关联研究
Beyond Owls: Subliminal Learning Can Transfer Learned Capabilities and Backdoors
在隐性学习中测试更多可传递能力与后门。
引用与复用
引用研究发现时请引用原始出版物；引用 NUWA 的整理与解释时，单独引用本站记录及对应快照。
原始出版物 · BibTeX
@misc{subliminal-learning,
title = {Subliminal Learning: Language models transmit behavioral traits via hidden signals in data},
url = {https://truthful.ai/papers/subliminal-learning/},
author = {Alex Cloud and Minh Le and James Chua and Jan Betley and Anna Sztyber-Betley and Sören Mindermann and Jacob Hilton and Samuel Marks and Owain Evans},
year = {2025},
date = {2025-07-20},
version = {Source capture 2026-10-25},
note = {Source capture 2026-10-25}
}
NUWA 记录 · BibTeX
@misc{subliminal-learning-nuwa,
title = {Subliminal Learning: Language models transmit behavioral traits via hidden signals in data — NUWA research record},
url = {https://whitzard.tech/nuwa/whitzard-index/work/subliminal-learning/},
version = {2026-10-29.1},
organization = {NUWA Frontier AI Safety Lab},
howpublished = {Retained note edition: https://whitzard.tech/nuwa/whitzard-index/data/record-editions/2026-10-29.1--subliminal-learning.json},
note = {NUWA research note. Source-reading scope and independent review status are stated on the record. Snapshot 2026-10-29.1. Retained note edition: https://whitzard.tech/nuwa/whitzard-index/data/record-editions/2026-10-29.1--subliminal-learning.json.}
}
证据快照 2026-10-29.1 · 索引更新 · RSS
参考文献
- [1] Alex Cloud, Minh Le 等. Subliminal Learning: Language models transmit behavioral traits via hidden signals in data. 2025-07-20 ; Official publication masthead and article byline; 2 Experimental setup: distillation on an unrelated domain; 3 Models transmit traits via numbers
- [2] Jan DubiÅski, Anna Sztyber-Betley 等. Beyond Owls: Subliminal Learning Can Transfer Learned Capabilities and Backdoors. 2026-10-07 ; Research background and discussion in the original publication