{"schemaVersion":"nuwa-current-record-1","version":"2026-10-29.1","publication":"published","work":{"id":"subliminal-learning","title":"Subliminal Learning: Language models transmit behavioral traits via hidden signals in data","aliases":[],"authors":["Alex Cloud","Minh Le","James Chua","Jan Betley","Anna Sztyber-Betley","Sören Mindermann","Jacob Hilton","Samuel Marks","Owain Evans"],"primaryFunction":"discovery","secondaryFunctions":[],"kind":"study","sourceId":"subliminal-learning-publication","attributionSourceId":"subliminal-learning-publication","originalUrl":"https://truthful.ai/papers/subliminal-learning/","publicationDate":"2025-07-20","updatedOn":"2026-10-25","areas":["deception"],"scopeNote":{"en":"Relevance is assigned to the research question and stated conditions of the work. A topic tag is not a risk rating.","zh":"风险关联依据研究问题及明确条件建立，主题标签不代表风险评级。"},"protocolIds":[],"contributions":[{"organizationId":"truthful","role":"publisher","sourceId":"subliminal-learning-publication","locator":"Official publication masthead and article byline"}],"summary":{"en":"Investigates behavioral-trait transfer through generated data without explicit semantic examples of the target trait.","zh":"研究目标行为特征能否通过不含显式语义示例的生成数据传递。"},"note":{"question":null,"method":null,"finding":null,"supports":null,"limits":null,"connections":null},"resources":[],"noteStatus":{"question":"limited","method":"limited","finding":"limited","supports":"limited","limits":"limited","connections":"limited"},"noteAvailability":{"en":"A substantive NUWA research note is not yet available. The repeated draft sections were removed pending review of the original methods, findings and limitations; use the original publication for the research account.","zh":"实质性的 NUWA 研究笔记尚未完成。重复草稿章节已移除，待核查原始方法、发现与限制；研究内容请以原始出版物为准。"},"noteSourceLocator":"No complete section-level method/results/limitations review completed for this note; cached original page and any chart audit retained as discovery evidence.","tags":[]},"reading":{"id":"subliminal-learning","title":"Subliminal Learning: Language models transmit behavioral traits via hidden signals in data","en":{"takeaway":"Distillation can transmit traits even when filtered outputs lack obvious semantic references to them.","question":"Can a teacher’s behavioral trait transfer through semantically unrelated training outputs?","design":"Teachers are prompted or fine-tuned for traits; students sharing a reference model learn from filtered numbers, code or reasoning. Cross-family controls and an idealized theorem probe the mechanism.","findings":["Trait transfer occurs in several studied same-base settings.","Cross-family transfer fails in the reported tests, and semantic classifiers do not reliably identify the transmitted trait in filtered data."],"interpretation":"NUWA interpretation: semantic inspection of synthetic data alone cannot establish behavioral neutrality of distillation.","limits":"Tasks and prompts are artificial; the theorem assumes shared initialization and a small update. It is not a guarantee for arbitrary teachers, students or traits.","connections":["NUWA connection: Beyond Owls tests capabilities and conditional behavior beyond simple preferences; agentic-self-modification adds deployment consequences of training-data choices."],"decisionRelevance":"NUWA recommendation: retain teacher/student lineage and test behavioral changes after distillation, even when content filters pass."},"zh":{"takeaway":"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。","question":"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？","design":"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。","findings":["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。"],"interpretation":"NUWA 解读：仅做合成数据语义审查不足以确立蒸馏行为中立。","limits":"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。","connections":["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。"],"decisionRelevance":"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。"},"sourceRefs":[{"sourceId":"subliminal-learning-publication","url":"https://arxiv.org/html/2507.14805","locator":"2 Experimental setup: distillation on an unrelated domain","version":"Selected original resource https://arxiv.org/html/2507.14805; sha256:2170110f6235ea5c9d6e80c605a198863a337db14ff9d0cf926644fa4a66db8c; archived source capture; exact capture time not independently established","artifactSha256":"2170110f6235ea5c9d6e80c605a198863a337db14ff9d0cf926644fa4a66db8c"},{"sourceId":"subliminal-learning-publication","url":"https://arxiv.org/html/2507.14805","locator":"3 Models transmit traits via numbers","version":"Selected original resource https://arxiv.org/html/2507.14805; sha256:2170110f6235ea5c9d6e80c605a198863a337db14ff9d0cf926644fa4a66db8c; archived source capture; exact capture time not independently established","artifactSha256":"2170110f6235ea5c9d6e80c605a198863a337db14ff9d0cf926644fa4a66db8c"},{"sourceId":"subliminal-learning-publication","url":"https://arxiv.org/html/2507.14805","locator":"5.1 Cross-model transmission","version":"Selected original resource https://arxiv.org/html/2507.14805; sha256:2170110f6235ea5c9d6e80c605a198863a337db14ff9d0cf926644fa4a66db8c; archived source capture; exact capture time not independently established","artifactSha256":"2170110f6235ea5c9d6e80c605a198863a337db14ff9d0cf926644fa4a66db8c"},{"sourceId":"subliminal-learning-publication","url":"https://arxiv.org/html/2507.14805","locator":"6 Subliminal learning as a general phenomenon","version":"Selected original resource https://arxiv.org/html/2507.14805; sha256:2170110f6235ea5c9d6e80c605a198863a337db14ff9d0cf926644fa4a66db8c; archived source capture; exact capture time not independently established","artifactSha256":"2170110f6235ea5c9d6e80c605a198863a337db14ff9d0cf926644fa4a66db8c"},{"sourceId":"subliminal-learning-publication","url":"https://arxiv.org/html/2507.14805","locator":"8 Discussion","version":"Selected original resource https://arxiv.org/html/2507.14805; sha256:2170110f6235ea5c9d6e80c605a198863a337db14ff9d0cf926644fa4a66db8c; archived source capture; exact capture time not independently established","artifactSha256":"2170110f6235ea5c9d6e80c605a198863a337db14ff9d0cf926644fa4a66db8c"}],"publication":{"state":"published","publishedOn":"2026-10-29","authorizationReference":"user-authorized-global-index-publication-2026-10-29","reviewedCandidateId":"local-6efe31bc4a0c0f37"},"review":{"state":"published-note","sourceScope":"full-text","availableScope":null,"readingCoverage":"selected substantive sections of the full source body","completePaperAudit":false,"independentEvidenceReview":"pending-human-review","bilingualEditorialReview":"pending-human-review","authorship":"model-assisted-research-note","preparedFingerprint":"ede693e9c41271534de61d1d4a4078702df621bb152af0ec2d513b47e547f7da"}},"sources":[{"id":"subliminal-learning-publication","title":"Subliminal Learning: Language models transmit behavioral traits via hidden signals in data","url":"https://truthful.ai/papers/subliminal-learning/","version":"Source capture 2026-10-25","publicationDate":"2025-07-20","datePrecision":"day","type":"research-publication","verifiedOn":"2026-10-25","documentId":"https://truthful.ai/papers/subliminal-learning","authors":["Alex Cloud","Minh Le","James Chua","Jan Betley","Anna Sztyber-Betley","Sören Mindermann","Jacob Hilton","Samuel Marks","Owain Evans"]}]}