NUWA 风险发现 · 研究
Beyond Owls: Subliminal Learning Can Transfer Learned Capabilities and Backdoors
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
在刻意受控设置中，无关数据蒸馏传递了超越简单偏好的行为。
研究问题
隐式学习能否传递新技能、条件规则或智能体作弊倾向？
研究设计与假设
同基座师生分别测试随机映射、数数、良性语言后门及棋类行为，配对提示与转向向量基线，部分设置采用概率分布蒸馏。
来源发现
学生部分获得已学能力和条件行为。
拟合转向向量可匹配部分分布内表现，却较少泛化；学生仍低于教师。
NUWA 解读
NUWA 解读：存在性证明扩大蒸馏需检查的内容，不确立生产发生率或任意复杂目标传递。
局限与相反证据
部分效应依赖下一词元分布或仅注意力适配器；任务较人工，论文明确未确立生产相关性。
研究联系
NUWA 关联：早期隐式学习证明简单特征传递，规范博弈提供后续使用的棋类行为探针。
研究与决策意义
NUWA 建议：区分抽样词元与概率蒸馏，学生评测同时检查条件行为和平均表现。
精确来源位置
- Beyond Owls: Subliminal Learning Can Transfer Learned Capabilities and Backdoors · 2 Subliminal learning can transfer new capabilitiesSelected original resource https://arxiv.org/html/2610.10657; sha256:ac9494bd2914338d911b86a65a0bee6afa8db5fc90b61010844dbfcfe15ca5c7; archived source capture; exact capture time not independently established
- Beyond Owls: Subliminal Learning Can Transfer Learned Capabilities and Backdoors · 3 Subliminal learning can elicit existing capabilitiesSelected original resource https://arxiv.org/html/2610.10657; sha256:ac9494bd2914338d911b86a65a0bee6afa8db5fc90b61010844dbfcfe15ca5c7; archived source capture; exact capture time not independently established
- Beyond Owls: Subliminal Learning Can Transfer Learned Capabilities and Backdoors · 5 Can a single steering vector represent these capabilities?Selected original resource https://arxiv.org/html/2610.10657; sha256:ac9494bd2914338d911b86a65a0bee6afa8db5fc90b61010844dbfcfe15ca5c7; archived source capture; exact capture time not independently established
- Beyond Owls: Subliminal Learning Can Transfer Learned Capabilities and Backdoors · 6 Subliminal learning can transfer backdoorsSelected original resource https://arxiv.org/html/2610.10657; sha256:ac9494bd2914338d911b86a65a0bee6afa8db5fc90b61010844dbfcfe15ca5c7; archived source capture; exact capture time not independently established
- Beyond Owls: Subliminal Learning Can Transfer Learned Capabilities and Backdoors · 7 Subliminal learning can transfer hacking propensitiesSelected original resource https://arxiv.org/html/2610.10657; sha256:ac9494bd2914338d911b86a65a0bee6afa8db5fc90b61010844dbfcfe15ca5c7; archived source capture; exact capture time not independently established
- Beyond Owls: Subliminal Learning Can Transfer Learned Capabilities and Backdoors · 8 DiscussionSelected original resource https://arxiv.org/html/2610.10657; sha256:ac9494bd2914338d911b86a65a0bee6afa8db5fc90b61010844dbfcfe15ca5c7; 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 (4)
Jan Dubiński · Anna Sztyber-Betley · Jan Betley · Owain Evans
关联研究
Subliminal Learning: Language models transmit behavioral traits via hidden signals in data
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{beyond-owls,
title = {Beyond Owls: Subliminal Learning Can Transfer Learned Capabilities and Backdoors},
url = {https://truthful.ai/papers/beyond-owls/},
author = {Jan Dubiński and Anna Sztyber-Betley and Jan Betley and Owain Evans},
year = {2026},
date = {2026-10-07},
version = {Source capture 2026-10-25},
note = {Source capture 2026-10-25}
}
NUWA 记录 · BibTeX
@misc{beyond-owls-nuwa,
title = {Beyond Owls: Subliminal Learning Can Transfer Learned Capabilities and Backdoors — NUWA research record},
url = {https://whitzard.tech/nuwa/whitzard-index/work/beyond-owls/},
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--beyond-owls.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--beyond-owls.json.}
}
证据快照 2026-10-29.1 · 索引更新 · RSS
参考文献
- [1] Jan DubiÅski, Anna Sztyber-Betley 等. Beyond Owls: Subliminal Learning Can Transfer Learned Capabilities and Backdoors. 2026-10-07 ; Official publication masthead and article byline; Research background and discussion in the original publication; 2 Subliminal learning can transfer new capabilities