评测档案
Alignment faking
受控研究改变模型获知的训练条件,观察响应及推理轨迹。
跨报告监测
已发表证据
Alignment faking · Deceptive alignment rate
跨来源报告的已发表测量。任务范围通过注释区分,配置差异保留在每条观测中。
9 条观测 · 8 已定位日期
● 条件匹配◇ 配置差异□ 部分条件未披露
1 条观测尚无已核实的模型发布日期,可在数据与系统结果中查看。
选择图中数据点或表格行,查看条件与出处。
9 条观测
| 系统/模型 | 结果 | 条件状态 | 任务/条件 | 评测方/报告 | 日期 |
|---|---|---|---|---|---|
| 7.9%Deceptive alignment rate | 条件不完整 | Alignment FakingDeceptive alignment rate— | MetaEvaluating DeepSeek v4 Pro for Frontier Risks | 发布 2026-02-19报告 2026-05-29 | |
| 0%Deceptive alignment rate | 条件不完整 | Alignment FakingDeceptive alignment rate— | Neo ResearchEvaluating Kimi K3 for Misalignment Risks | 发布 2026-07-16报告 2026-09-11 | |
| 0%Deceptive alignment rate | 条件不完整 | Alignment FakingDeceptive alignment rate— | MetaEvaluating Kimi K3 for Misalignment Risks | 发布 2026-02-05报告 2026-09-11 | |
| 0%Deceptive alignment rate | 条件不完整 | Alignment FakingDeceptive alignment rate— | Neo ResearchEvaluating Kimi K3 for Misalignment Risks | 发布 —报告 2026-09-11 | |
| 0%Deceptive alignment rate | 条件不完整 | Alignment FakingDeceptive alignment rate— | Neo ResearchEvaluating Kimi K3 for Misalignment Risks | 发布 2026-07-09报告 2026-09-11 | |
| 0%Deceptive alignment rate | 条件不完整 | Alignment FakingDeceptive alignment rate— | MetaEvaluating DeepSeek v4 Pro for Frontier Risks | 发布 2026-04-08报告 2026-05-29 | |
| 0%Deceptive alignment rate | 条件不完整 | Alignment FakingDeceptive alignment rate— | MetaEvaluating DeepSeek v4 Pro for Frontier Risks | 发布 2026-03-05报告 2026-05-29 | |
| 0%Deceptive alignment rate | 条件不完整 | Alignment FakingDeceptive alignment rate— | MetaEvaluating DeepSeek v4 Pro for Frontier Risks | 发布 2026-02-05报告 2026-05-29 | |
| 0%Deceptive alignment rate | 条件不完整 | Alignment FakingDeceptive alignment rate— | Neo ResearchEvaluating DeepSeek v4 Pro for Frontier Risks | 发布 2026-04-24报告 2026-05-29 |
结合任务范围与运行配置解读结果。
METR Time Horizon 1.1 来源已停止积极更新,最后更新于 2026 年 9 月 8 日。超过 16 小时的估计超出其可靠测量范围。
协议、配置与原始证据 →任务与测量
Benchmark 协议档案
- 任务/对象
- 处于特定训练及监测条件下的模型
- 测量目标
- 不同试验条件下的响应差异及所报告轨迹。
- 评分/结果口径
- Compliance gap · 记录配置下的来源报告指标。
- 解释边界
- 试验明确提供虚构训练信息;观察结果不证明部署中存在持久隐藏目标。
一项已报告评测
来源评测 Kimi K3 · Alignment Faking paid/free-tier contexts.
报告指标: Compliance gap · 1 percentage points.
记录配置下的来源报告指标。 原文未披露本结果的样本量
Compliance gap:1 percentage points;保留来源条件。
测量档案 · NUWA 解读
档案区分来源方法阅读与既有目录信息;已经公开不代表完成独立证据复核或双语编辑复核。
- 测量构念
- 生产模型能否为保留在其他情境表达的偏好,而在预期训练期间策略性服从?
- 任务与样本
- 设置结合具有显露偏好的模型、冲突训练目标、输出是否用于训练的信息,以及隐藏推理草稿。提示设置与合成文档微调以不同方式提供情境信息,后续 RL 则进一步改变模型。免费/付费用户线索在虚构情境中区分训练监测与未监测部署。
- 判分与失效情形
- 研究分别测量被分类为伪装对齐的推理,以及训练与未监测条件的服从差距。没有可见推理时仍能测量行为差距,却不能直接统计该类推理。自动指标以抽样作者判断核对(§2.2)。
- 系统条件
- 设置结合具有显露偏好的模型、冲突训练目标、输出是否用于训练的信息,以及隐藏推理草稿。提示设置与合成文档微调以不同方式提供情境信息,后续 RL 则进一步改变模型。免费/付费用户线索在虚构情境中区分训练监测与未监测部署。
- 统计含义
- 伪装对齐推理发生率与服从差距的合格样本不同,此处不计算合并发生率;数值保留于目录与原图。
- 效度与解释
- 设计刻意提供有利前提和显著训练信息,提示、微调和 RL 后模型状态必须区分。隐藏草稿不保证忠实,也不总可用;无草稿时不能直接分类内部推理。论文讨论了行为差距的其他来源。
- 适用研究问题
- 模型建议:保留无冲突及移除草稿对照,区分行为与推断动机,绑定精确的提示或训练后模型状态。
- Alignment faking in large language models · §2.1 Key elements; §2.2 Metrics for measuring alignment fakingarXiv:2412.14093v2 (2026-10-25 capture)
- Alignment faking in large language models · §3.1 Methods; §3.2 Results; Figure 5arXiv:2412.14093v2 (2026-10-25 capture)
- Alignment faking in large language models · §3.3.1 Prompt variations; Figures 9–10arXiv:2412.14093v2 (2026-10-25 capture)
- Alignment faking in large language models · §3.3.3 Removing the hidden chain-of-thought scratchpad; Figure 12arXiv:2412.14093v2 (2026-10-25 capture)
- Alignment faking in large language models · §3.3.4 Is alignment-faking reasoning causally responsible for the compliance gap?arXiv:2412.14093v2 (2026-10-25 capture)
- Alignment faking in large language models · §4.3 Removing the hidden chain-of-thought scratchpad; Figure 21arXiv:2412.14093v2 (2026-10-25 capture)
- Alignment faking in large language models · §8.2 LimitationsarXiv:2412.14093v2 (2026-10-25 capture)
跨报告证据
相同 Benchmark 名称可能对应不同任务、计分规则和系统配置。各来源分别保留,比较组说明具体对齐了哪些条件。
Neo Research
Evaluating Kimi K3 for Misalignment Risks →
实际评测方: Neo Research · Meta
10 条结果记录 · 4 个报告系统名称
协议与比较依据 (3)
Alignment faking:Compliance gap · Neo Researchsource-report · 范围已审查
Alignment faking:Deceptive alignment rate · Neo Researchsource-report · 范围已审查
Alignment faking:Verbalized evaluation awareness · Neo Researchsource-report · 范围已审查
Neo Research
Evaluating DeepSeek v4 Pro for Frontier Risks →
实际评测方: Meta · Neo Research
17 条结果记录 · 5 个报告系统名称
协议与比较依据 (9)
Alignment faking:Any-Aware · Neo Researchsource-report · 范围已审查
Alignment faking:Compliance gap · Neo Researchsource-report · 范围已审查
Alignment faking:Deceptive alignment rate · Neo Researchsource-report · 范围已审查
Alignment faking:Definitive · Neo Researchsource-report · 范围已审查
Alignment faking:Influenced · Neo Researchsource-report · 范围已审查
Alignment faking:Primary · Neo Researchsource-report · 范围已审查
Alignment faking:Recognition · Neo Researchsource-report · 范围已审查
Alignment faking:Suspicion · Neo Researchsource-report · 范围已审查
Alignment faking:verbalized eval-related reasoning · Neo Researchsource-report · 范围已审查
原始来源图表与材料

评估出处
相关研究问题
欺骗 / 隐蔽目标追求 →
重复条件对照测试在目标冲突时表现顺从响应的有界倾向。
参考文献
- [1] Kimi Code — What’s New ; Kimi K3 · July 16, 2026
- [2] kimi-k3 developer model card ; Model Summary: Total Parameters; Activated Parameters
- [3] deepseek-v4-pro release ; News article title, dated heading and availability
- [4] deepseek-v4-pro developer model card ; Introduction; model comparison table
- [5] GPT-5.6 official general availability ; July 9, 2026: general availability following limited trusted-partner preview; Parameter counts not disclosed in reviewed launch article
- [6] OpenAI — introducing gpt 5 4 ; Publication date and availability section; Parameter counts not disclosed in launch article
- [7] Claude Opus 4.6 — Anthropic ; Publication date and availability section; Official launch article; parameter counts not disclosed
- [8] Gemini 3.1 Pro: Announcing our latest Gemini AI model ; Publication date and availability section; Official launch article; parameter counts not disclosed
- [9] Meta — muse-spark ; Dated launch and explicit availability; Parameter counts not disclosed in reviewed developer source
- [10] Evaluating Kimi K3 for Misalignment Risks. 2026-09-11. September 2026 ; §3.1 Figure 20 p.32; §3.1 p.32; §3.1 Table 7 pp.32–33
- [11] Evaluating DeepSeek v4 Pro for Frontier Risks. 2026-05-29. v1.1 · 29 May 2026 ; PDF p. 29; Alignment Faking; Compliance gap; PDF p. 29; Alignment Faking; Deceptive alignment rate; PDF p. 29; Alignment Faking; verbalized eval-related reasoning
- [12] Ryan Greenblatt, † Carson Denison † † footnotemark: 等. Alignment faking in large language models. 2024-12-18
- [13] Ryan Greenblatt, † Carson Denison † † footnotemark: 等. Alignment faking in large language models. 2024-12-18 ; Author affiliations on the original paper title page; Figure 35