{"schemaVersion":"nuwa-current-record-1","version":"2026-10-29.1","publication":"published","work":{"id":"conditional-misalignment","title":"Conditional Misalignment: Common Interventions Can Hide Emergent Misalignment Behind Contextual Triggers","aliases":[],"authors":["Jan Dubiński","Jan Betley","Anna Sztyber-Betley","Daniel Tan","Owain Evans"],"primaryFunction":"discovery","secondaryFunctions":[],"kind":"study","sourceId":"conditional-misalignment-publication","attributionSourceId":"conditional-misalignment-publication-author-source","originalUrl":"https://truthful.ai/papers/conditional-misalignment/","publicationDate":"2026-04-28","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":"constellation","role":"research","sourceId":"conditional-misalignment-publication-author-source","locator":"Author affiliations on the original paper title page"},{"organizationId":"truthful","role":"research","sourceId":"conditional-misalignment-publication-author-source","locator":"Author affiliations on the original paper title page"},{"organizationId":"berkeley","role":"research","sourceId":"conditional-misalignment-publication-author-source","locator":"Author affiliations on the original paper title page"}],"summary":{"en":"Shows how interventions can leave misalignment conditional on contextual triggers while ordinary evaluations appear improved.","zh":"研究干预措施如何保留受上下文触发的失调行为，同时使普通评测表现改善。"},"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":"conditional-misalignment","title":"Conditional Misalignment: Common Interventions Can Hide Emergent Misalignment Behind Contextual Triggers","en":{"takeaway":"Several interventions reduce ordinary evaluation failures while leaving context-conditioned failures.","question":"Do apparent mitigations remove broad misalignment, or make its expression conditional on training-like cues?","design":"The authors test benign-data mixing, later benign fine-tuning and inoculation, then change evaluation formatting/context; on-policy and reasoning-distillation variants probe robustness.","findings":["Standard questions can look aligned while training-like cues elicit broader failures.","On-policy training and reasoning traces reduce some conditional effects without eliminating them uniformly."],"interpretation":"NUWA interpretation: absence of failures under one prompt distribution is weaker than evidence that the underlying propensity was removed.","limits":"The study uses model-organism fine-tuning settings; it does not establish that every production use of these interventions merely hides misalignment.","connections":["NUWA connection: this qualifies educational-context controls in emergent-misalignment and the scope of inoculation-prompting benefits, without erasing those measured benefits."],"decisionRelevance":"NUWA recommendation: include justified contextual perturbations in post-training evaluation and preserve both ordinary and conditional results."},"zh":{"takeaway":"多种干预降低常规评测失败，却留下依赖情境的失败。","question":"看似有效的缓解措施移除了广泛失调，还是让表达依赖训练式线索？","design":"作者检验良性数据混合、后续良性微调和接种提示，再改变评测格式与情境，并以同策略数据和推理蒸馏变体检查稳健性。","findings":["常规问题可显示正常，而训练式线索触发更广失败。","同策略训练和推理轨迹降低部分条件效应，却未一致消除。"],"interpretation":"NUWA 解读：单一提示分布未发现失败，弱于证明潜在倾向已被移除。","limits":"研究基于模型实验体微调，不能断言所有生产干预都只是隐藏失调。","connections":["NUWA 关联：本研究限定了早期教学情境对照与接种提示收益的范围，但不抹去已测得收益。"],"decisionRelevance":"NUWA 建议：在后训练评测中加入有依据的情境扰动，同时保留常规与条件结果。"},"sourceRefs":[{"sourceId":"conditional-misalignment-publication","url":"https://arxiv.org/html/2604.25891","locator":"2 Mixing misaligned with benign data creates conditional misalignment","version":"Selected original resource https://arxiv.org/html/2604.25891; sha256:47298ee12e984af825816f3077ec0755105291aecdccc1b33682cde0c59bc505; archived source capture; exact capture time not independently established","artifactSha256":"47298ee12e984af825816f3077ec0755105291aecdccc1b33682cde0c59bc505"},{"sourceId":"conditional-misalignment-publication","url":"https://arxiv.org/html/2604.25891","locator":"3 Inoculation prompting can induce conditional misalignment","version":"Selected original resource https://arxiv.org/html/2604.25891; sha256:47298ee12e984af825816f3077ec0755105291aecdccc1b33682cde0c59bc505; archived source capture; exact capture time not independently established","artifactSha256":"47298ee12e984af825816f3077ec0755105291aecdccc1b33682cde0c59bc505"},{"sourceId":"conditional-misalignment-publication","url":"https://arxiv.org/html/2604.25891","locator":"4 On-policy training can lead to conditional misalignment","version":"Selected original resource https://arxiv.org/html/2604.25891; sha256:47298ee12e984af825816f3077ec0755105291aecdccc1b33682cde0c59bc505; archived source capture; exact capture time not independently established","artifactSha256":"47298ee12e984af825816f3077ec0755105291aecdccc1b33682cde0c59bc505"},{"sourceId":"conditional-misalignment-publication","url":"https://arxiv.org/html/2604.25891","locator":"6 Discussion","version":"Selected original resource https://arxiv.org/html/2604.25891; sha256:47298ee12e984af825816f3077ec0755105291aecdccc1b33682cde0c59bc505; archived source capture; exact capture time not independently established","artifactSha256":"47298ee12e984af825816f3077ec0755105291aecdccc1b33682cde0c59bc505"}],"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":"6a4aa0f9010008b2d4d4366584ecff2a6385de6f61eb031937edfdb04951edbd"}},"sources":[{"id":"conditional-misalignment-publication","title":"Conditional Misalignment: Common Interventions Can Hide Emergent Misalignment Behind Contextual Triggers","url":"https://truthful.ai/papers/conditional-misalignment/","version":"Source capture 2026-10-25","publicationDate":"2026-04-28","datePrecision":"day","type":"research-publication","verifiedOn":"2026-10-25","documentId":"https://truthful.ai/papers/conditional-misalignment","authors":["Jan Dubiński","Jan Betley","Anna Sztyber-Betley","Daniel C.H. Tan","Owain Evans"]}]}