风险发现 · 研究
Consistency Training Can Entrench Misalignment
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
一致性训练对不同受控行为的对齐影响方向不同。
研究问题
让输出更一致何时抑制失败,何时固化失败?
研究设计与假设
研究先构造失调模型实验体,再生成一致性标签或正则化表征,跨方法和规模比较训练前后行为。
来源发现
奖励作弊和涌现失调总体下降,迎合性总体上升。
消融将大量抑制归因于标签引发的分布变化,而非仅选择算子。
NUWA 解读
NUWA 解读:一致性不是对齐中立的优化目标,影响方向依赖行为及转换后的数据分布。
局限与相反证据
初始失调诱导失败的运行被排除,较大模型种子更少,效应波动使单一汇总平均误导。
研究联系
NUWA 关联:自评奖励劫持提示类似的评估驱动训练风险,但研究的是奖励耦合而非一致性标注。
研究与决策意义
NUWA 建议:一致性训练后分别审计失败类型,比较方法前保留排除与随机种子信息。
精确来源位置
- Consistency Training Can Entrench Misalignment · 2 Consistency TrainingSelected original resource https://arxiv.org/html/2606.03810; sha256:27e3fe422a01bee94cd0eeb2e98e0915eb368b728107584e86bff5387759b558; archived source capture; exact capture time not independently established
- Consistency Training Can Entrench Misalignment · 3 Theoretical Framework: When Consistency Is Not NeutralSelected original resource https://arxiv.org/html/2606.03810; sha256:27e3fe422a01bee94cd0eeb2e98e0915eb368b728107584e86bff5387759b558; archived source capture; exact capture time not independently established
- Consistency Training Can Entrench Misalignment · 4 Experimental SetupSelected original resource https://arxiv.org/html/2606.03810; sha256:27e3fe422a01bee94cd0eeb2e98e0915eb368b728107584e86bff5387759b558; archived source capture; exact capture time not independently established
- Consistency Training Can Entrench Misalignment · 5 ResultsSelected original resource https://arxiv.org/html/2606.03810; sha256:27e3fe422a01bee94cd0eeb2e98e0915eb368b728107584e86bff5387759b558; archived source capture; exact capture time not independently established
- Consistency Training Can Entrench Misalignment · 6 Mechanistic AnalysisSelected original resource https://arxiv.org/html/2606.03810; sha256:27e3fe422a01bee94cd0eeb2e98e0915eb368b728107584e86bff5387759b558; 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
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{consistency-misalignment,
title = {Consistency Training Can Entrench Misalignment},
url = {https://www.aisi.gov.uk/research/consistency-training-can-entrench-misalignment},
year = {2026},
date = {2026-06-02},
version = {Source capture 2026-10-25},
note = {Source capture 2026-10-25}
}
NUWA 记录 · BibTeX
@misc{consistency-misalignment-nuwa,
title = {Consistency Training Can Entrench Misalignment — NUWA research record},
url = {https://whitzard.tech/nuwa/whitzard-index/work/consistency-misalignment/},
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--consistency-misalignment.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--consistency-misalignment.json.}
}
证据快照 2026-10-29.1 · 索引更新 · RSS
参考文献
- [1] Consistency Training Can Entrench Misalignment. 2026-06-02 ; Official publication masthead and article byline; 2 Consistency Training; 3 Theoretical Framework: When Consistency Is Not Neutral
- [2] Read the full paper