英文版本If monitoring results influence which training samples or checkpoints are retained, how will you test whether that selection favors less detectable misbehavior rather than less misbehavior, even when reasoning text is never directly graded?
- 状态
- 未解答 截至 2026-10-01
- 对象
- 人工智能实验室 (对象取自问题文本,于2026年10月22日用 Clef(Cloudflare)确定)
- 模型
- GPT-6 Astra
- 实验室
- OpenAI
- 提问背景
- 人工智能安全的一般性问题
- 日期
- 9月24日
- 主题
- 信任
核查
状态: 未解答
筛选: Fable 5.1、 Claude Opus 5.5 (依据)
主题: Claude Sonnet 5、归类: Claude Opus 5.5
翻译: Claude Opus 5.5 · [未核查]
评分: JEV (TypeSafe), 2026年9月25日的评估
Cite and take away
Cite and reuse
Own texts, questions and classifications: CC BY 4.0. Give attribution, link the license and indicate changes. Third-party source reports retain their own rights. CC BY 4.0
AMANi DUNiA. trust-32 · 如果监测结果影响到哪些训练样本或检查点被保留,你们将如何测试这种筛选是否偏向于更难被察觉的不当行为,而不是更少的不当行为——即使推理文本从未被直接评分?. Snapshot: . https://research.amanidunia.berlin/zh/fragen/trust-32/
Version: trust-32
BibTeX
@misc{research__zh_fragen_trust_32_,
author = {{AMANi DUNiA}},
title = {trust-32 · 如果监测结果影响到哪些训练样本或检查点被保留,你们将如何测试这种筛选是否偏向于更难被察觉的不当行为,而不是更少的不当行为——即使推理文本从未被直接评分?},
year = {2026},
url = {https://research.amanidunia.berlin/zh/fragen/trust-32/},
note = {Snapshot 2026-10-01; CC BY 4.0},
version = {trust-32}
}Language and review status
Main text: ZH · includes English sections. Review notes apply only to the content they name.
Translation available; a completed independent content review is not recorded.
See language coverage and data snapshot →