风险治理 · 技术报告

We need 3rd party Training-Run Evaluations

NUWA 研究解读 · 独立复核待完成

首次发表: 2026-07-05来源核验: 2026-10-25

原始来源 ↗引用本记录 ↓

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

该提案将部分安全评估从最终检查点移至训练动态及开发者响应。

研究问题

哪些训练期证据能揭示最终检查点评测可能漏掉的持续图谋？

研究设计与假设

文章分析隐蔽行为与难读认知，区分中间检查点访问，以及对数据、奖励、环境和干预的更深入访问。

来源发现

评估者直接访问检查点可支持自适应调查，超出固定的客户代跑测试。

作者区分分散与集中的知识产权风险，建议逐步增加访问，并未报告通用 TRE 标准已完成。

NUWA 解读

NUWA 解读：没有过程证据，最终测试阴性不能验证整个训练过程的安全主张。

局限与相反证据

这是明确承认图谋可检测性不确定的机构论证，不是 TRE 能发现所有图谋的实验证明。

研究联系

NUWA 关联：嵌入式评估将访问论证扩展到持续部署与治理；反图谋训练提供具体评估意识疑虑。

研究与决策意义

NUWA 建议：外部保证前，将训练安全主张逐项绑定所需访问、检查点时机和未提供的信息。

精确来源位置

We need 3rd party Training-Run Evaluations · Detecting Scheming may require Training-Run EvaluationsSelected original resource https://www.apolloresearch.ai/science/we-need-3rd-party-training-run-assessments; sha256:c4d034daec43cf8030397f4b36514eb1a4d44ac2b1037ee97a15a8ddbad116a0; archived source capture; exact capture time not independently established
We need 3rd party Training-Run Evaluations · Why 3rd parties should perform Training-Run EvaluationsSelected original resource https://www.apolloresearch.ai/science/we-need-3rd-party-training-run-assessments; sha256:c4d034daec43cf8030397f4b36514eb1a4d44ac2b1037ee97a15a8ddbad116a0; archived source capture; exact capture time not independently established
We need 3rd party Training-Run Evaluations · Checkpoint evalsSelected original resource https://www.apolloresearch.ai/science/we-need-3rd-party-training-run-assessments; sha256:c4d034daec43cf8030397f4b36514eb1a4d44ac2b1037ee97a15a8ddbad116a0; archived source capture; exact capture time not independently established
We need 3rd party Training-Run Evaluations · Data EvaluationsSelected original resource https://www.apolloresearch.ai/science/we-need-3rd-party-training-run-assessments; sha256:c4d034daec43cf8030397f4b36514eb1a4d44ac2b1037ee97a15a8ddbad116a0; archived source capture; exact capture time not independently established
We need 3rd party Training-Run Evaluations · ConclusionSelected original resource https://www.apolloresearch.ai/science/we-need-3rd-party-training-run-assessments; sha256:c4d034daec43cf8030397f4b36514eb1a4d44ac2b1037ee97a15a8ddbad116a0; 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

发布机构 Apollo Research

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{training-run-evaluations,
  title = {We need 3rd party Training-Run Evaluations},
  url = {https://www.apolloresearch.ai/science/we-need-3rd-party-training-run-assessments},
  year = {2026},
  date = {2026-07-05},
  version = {Source capture 2026-10-25},
  note = {Source capture 2026-10-25}
}

NUWA 记录 · BibTeX

@misc{training-run-evaluations-nuwa,
  title = {We need 3rd party Training-Run Evaluations — NUWA research record},
  url = {https://whitzard.tech/nuwa/whitzard-index/work/training-run-evaluations/},
  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--training-run-evaluations.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--training-run-evaluations.json.}
}

证据快照 2026-10-29.1 · 索引更新 · RSS

参考文献

[1] We need 3rd party Training-Run Evaluations. 2026-07-05 ; Official publication masthead and article byline; Detecting Scheming may require Training-Run Evaluations; Why 3rd parties should perform Training-Run Evaluations