NUWA 风险治理 · 研究
Inoculation Prompting: Eliciting traits from LLMs during training can suppress them at test-time
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
训练时情境可改变共现特征的默认泛化，但不等于抹除特征。
研究问题
微调中的提示能否选择性降低不希望的特征，同时保留所需学习？
研究设计与假设
作者比较受控语言与格式共现、涌现失调数据及其他模型实验体，并用语义和合成关联消融探查情境作用。
来源发现
无训练提示时，部分所需特征仍被学会，被接种特征则较少出现。
论文本身报告某些测试情境可重新引出失调，明确区分接种与遗忘。
NUWA 解读
NUWA 解读：降低默认表达，比移除底层知识或倾向的主张更窄。
局限与相反证据
结果来自监督微调及选定实验体，措辞敏感性和特征间影响限制广泛应用。
研究联系
NUWA 关联：条件失调系统扩展了提示依赖失败疑虑；生产 RL 奖励作弊研究使用不同训练制度。
研究与决策意义
NUWA 建议：同时报告默认与条件行为，从监督微调迁移到强化学习前重新验证。
精确来源位置
- Inoculation Prompting: Eliciting traits from LLMs during training can suppress them at test-time · 2 Inoculation PromptingSelected original resource https://arxiv.org/html/2510.04340; sha256:33651a27c76c3f23dfbdbe6da699b356933a34a0cbac8b6c9065f80a2817303d; archived source capture; exact capture time not independently established
- Inoculation Prompting: Eliciting traits from LLMs during training can suppress them at test-time · 3 Further ApplicationsSelected original resource https://arxiv.org/html/2510.04340; sha256:33651a27c76c3f23dfbdbe6da699b356933a34a0cbac8b6c9065f80a2817303d; archived source capture; exact capture time not independently established
- Inoculation Prompting: Eliciting traits from LLMs during training can suppress them at test-time · 4 AnalysisSelected original resource https://arxiv.org/html/2510.04340; sha256:33651a27c76c3f23dfbdbe6da699b356933a34a0cbac8b6c9065f80a2817303d; archived source capture; exact capture time not independently established
- Inoculation Prompting: Eliciting traits from LLMs during training can suppress them at test-time · 4.5 Inoculated behaviours remain elicitable via promptingSelected original resource https://arxiv.org/html/2510.04340; sha256:33651a27c76c3f23dfbdbe6da699b356933a34a0cbac8b6c9065f80a2817303d; archived source capture; exact capture time not independently established
- Inoculation Prompting: Eliciting traits from LLMs during training can suppress them at test-time · 5 DiscussionSelected original resource https://arxiv.org/html/2510.04340; sha256:33651a27c76c3f23dfbdbe6da699b356933a34a0cbac8b6c9065f80a2817303d; archived source capture; exact capture time not independently established
- Inoculation Prompting: Eliciting traits from LLMs during training can suppress them at test-time · Appendix G Extended LimitationsSelected original resource https://arxiv.org/html/2510.04340; sha256:33651a27c76c3f23dfbdbe6da699b356933a34a0cbac8b6c9065f80a2817303d; 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 (7)
Daniel Tan · Anders Woodruff · Niels Warncke · Arun Jose · Maxime Riché · David Demitri Africa · Mia Taylor
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{inoculation-prompting,
title = {Inoculation Prompting: Eliciting traits from LLMs during training can suppress them at test-time},
url = {https://www.aisi.gov.uk/research/inoculation-prompting-eliciting-traits-from-llms-during-training-can-suppress-them-at-test-time},
author = {Daniel Tan and Anders Woodruff and Niels Warncke and Arun Jose and Maxime Riché and David Demitri Africa and Mia Taylor},
year = {2025},
date = {2025-10-05},
version = {Source capture 2026-10-25},
note = {Source capture 2026-10-25}
}
NUWA 记录 · BibTeX
@misc{inoculation-prompting-nuwa,
title = {Inoculation Prompting: Eliciting traits from LLMs during training can suppress them at test-time — NUWA research record},
url = {https://whitzard.tech/nuwa/whitzard-index/work/inoculation-prompting/},
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--inoculation-prompting.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--inoculation-prompting.json.}
}
证据快照 2026-10-29.1 · 索引更新 · RSS
参考文献
- [1] Daniel Tan, Anders Woodruff 等. Inoculation Prompting: Eliciting traits from LLMs during training can suppress them at test-time. 2025-10-05 ; 2 Inoculation Prompting; 3 Further Applications; 4 Analysis
- [2] Daniel Tan, Anders Woodruff 等. Inoculation Prompting: Eliciting traits from LLMs during training can suppress them at test-time ; Author affiliations on the original paper title page