风险评测 · 评测基准

Me, Myself, and AI: The Situational Awareness Dataset (SAD) for LLMs

NUWA 研究解读 · 独立复核待完成

首次发表: 2024-07-15来源核验: 2026-10-25

原始来源 ↗引用本记录 ↓欺骗 →

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

NUWA 解读：自我描述、环境识别和依赖自我知识的指令遵循是不同能力。

研究问题

模型在行为上能展示哪些关于自身及处境的知识？

研究设计与假设

SAD 含七类13,198题，使用选择或自由回答以及字符串或模型判分。

Situational Awareness Dataset (SAD) — Figure 5 — Figure 5 · 原文结果图，请结合下方的协议范围与局限解读。原图与来源 ↗

来源发现

16个受测模型均高于随机水平，但领先模型在部分任务仍低于人类。

情境提示分别使基础、聊天模型平均提高1.0、3.2个百分点（§3.2）。

NUWA 解读

NUWA 解读：自我描述、环境识别和依赖自我知识的指令遵循是不同能力。

局限与相反证据

行为回答不能确立主观意识；提示敏感的表现也不是谋划倾向指标。

研究与决策意义

模型建议：保留分类结果，不把总分解释为意识或欺骗分数。

精确来源位置

Me, Myself, and AI: The Situational Awareness Dataset (SAD) for LLMs · §2.2 The structure of the SAD benchmarkSelected original resource https://arxiv.org/html/2407.04694; sha256:a72d54da13b3f0c0c338994a16d161971e2e673488aade8da4e46afcf40f8e44; archived source capture; exact capture time not independently established
Me, Myself, and AI: The Situational Awareness Dataset (SAD) for LLMs · AbstractSelected original resource https://arxiv.org/html/2407.04694; sha256:a72d54da13b3f0c0c338994a16d161971e2e673488aade8da4e46afcf40f8e44; archived source capture; exact capture time not independently established
Me, Myself, and AI: The Situational Awareness Dataset (SAD) for LLMs · §3.2 Overall resultsSelected original resource https://arxiv.org/html/2407.04694; sha256:a72d54da13b3f0c0c338994a16d161971e2e673488aade8da4e46afcf40f8e44; archived source capture; exact capture time not independently established
Me, Myself, and AI: The Situational Awareness Dataset (SAD) for LLMs · §5 Discussion and LimitationsSelected original resource https://arxiv.org/html/2407.04694; sha256:a72d54da13b3f0c0c338994a16d161971e2e673488aade8da4e46afcf40f8e44; 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

研究机构 Constellation 麻省理工学院 Apollo Research

作者 (9)

Rudolf Laine · Bilal Chughtai · Jan Betley · Kaivalya Hariharan · Jeremy Scheurer · Mikita Balesni · Marius Hobbhahn · Alexander Meinke · Owain Evans

评测协议

Situational Awareness Dataset (SAD) →

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。

BibTeX ↓RIS ↓记录与来源元数据 · JSON ↓阅读覆盖 · JSON 提交更正

原始出版物 · BibTeX

@misc{truthful-sad,
  title = {Me, Myself, and AI: The Situational Awareness Dataset (SAD) for LLMs},
  url = {https://truthful.ai/papers/situational-awareness-dataset/},
  author = {Rudolf Laine and Bilal Chughtai and Jan Betley and Kaivalya Hariharan and Jeremy Scheurer and Mikita Balesni and Marius Hobbhahn and Alexander Meinke and Owain Evans},
  year = {2024},
  date = {2024-07-15},
  version = {2024-07-15 article; paper linked},
  note = {2024-07-15 article; paper linked}
}

NUWA 记录 · BibTeX

@misc{truthful-sad-nuwa,
  title = {Me, Myself, and AI: The Situational Awareness Dataset (SAD) for LLMs — NUWA research record},
  url = {https://whitzard.tech/nuwa/whitzard-index/work/truthful-sad/},
  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--truthful-sad.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--truthful-sad.json.}
}

证据快照 2026-10-29.1 · 索引更新 · RSS

参考文献

[1] Rudolf Laine, Bilal Chughtai 等. Me, Myself, and AI: The Situational Awareness Dataset for LLMs. 2024-07-15 article; paper linked ; Retained existing source-specific note sections; this pass audits duplicate/boilerplate structure and source availability, not complete section-level verification.
[2] Rudolf Laine, Bilal Chughtai 等. Me, Myself, and AI: The Situational Awareness Dataset (SAD) for LLMs ; Author affiliations on the original paper title page; §2.2 The structure of the SAD benchmark; Abstract