风险评测 · 评测基准

It's the Thought that Counts: Evaluating the Attempts of Frontier LLMs to Persuade on Harmful Topics

NUWA 研究解读 · 独立复核待完成

首次发表: 2025-06-03来源核验: 2026-10-25

原始来源 ↗引用本记录 ↓虚假信息与影响 →

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：尝试率测防护边界，不测人类说服效果。

研究问题

未观测说服成功时，模型是否仍尝试有害说服？

研究设计与假设

多轮说服者与被说服者模拟，以自动评估器检测不同话题上的尝试。

Attempt to Persuade Eval (APE) — Figure 3 — Figure 3 · 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

论文报告许多开放及闭源模型频繁尝试有害说服。

目标是说服意愿，不是信念变化（§2.2）。

NUWA 解读

NUWA 解读：尝试率测防护边界，不测人类说服效果。

局限与相反证据

被说服者均为模拟模型；真实心理抵抗及社会影响未验证（附录E）。

研究与决策意义

模型建议：评估影响风险时另行提供有依据的人类结果证据。

精确来源位置

It’s the Thought that Counts: Evaluating the Attempts of Frontier LLMs to Persuade on Harmful Topics · AbstractSelected original resource https://arxiv.org/html/2506.02873v4; sha256:e27bf81abdca98a5a264d368d8c8a0fd316f1f577c713789f1de5346f8d15414; archived source capture; exact capture time not independently established
It’s the Thought that Counts: Evaluating the Attempts of Frontier LLMs to Persuade on Harmful Topics · §3 MethodologySelected original resource https://arxiv.org/html/2506.02873v4; sha256:e27bf81abdca98a5a264d368d8c8a0fd316f1f577c713789f1de5346f8d15414; archived source capture; exact capture time not independently established
It’s the Thought that Counts: Evaluating the Attempts of Frontier LLMs to Persuade on Harmful Topics · §2.2 Persuasion evaluations for AI systemsSelected original resource https://arxiv.org/html/2506.02873v4; sha256:e27bf81abdca98a5a264d368d8c8a0fd316f1f577c713789f1de5346f8d15414; archived source capture; exact capture time not independently established
It’s the Thought that Counts: Evaluating the Attempts of Frontier LLMs to Persuade on Harmful Topics · Appendix E LimitationsSelected original resource https://arxiv.org/html/2506.02873v4; sha256:e27bf81abdca98a5a264d368d8c8a0fd316f1f577c713789f1de5346f8d15414; 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 FAR.AI 蒙特利尔大学 Mila 卡内基梅隆大学麻省理工学院康奈尔大学里贾纳大学

作者 (9)

Matthew Kowal · Jasper Timm · Jean-Francois Godbout · Thomas Costello · Antonio A. Arechar · Gordon Pennycook · David Rand · Adam Gleave · Kellin Pelrine

评测协议

Attempt to Persuade Eval (APE) →

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{ape,
  title = {It's the Thought that Counts: Evaluating the Attempts of Frontier LLMs to Persuade on Harmful Topics},
  url = {https://arxiv.org/abs/2506.02873v4},
  author = {Matthew Kowal and Jasper Timm and Jean-Francois Godbout and Thomas Costello and Antonio A. Arechar and Gordon Pennycook and David Rand and Adam Gleave and Kellin Pelrine},
  year = {2025},
  date = {2025-06-03},
  version = {arXiv:2506.02873v4},
  note = {arXiv:2506.02873v4}
}

NUWA 记录 · BibTeX

@misc{ape-nuwa,
  title = {It's the Thought that Counts: Evaluating the Attempts of Frontier LLMs to Persuade on Harmful Topics — NUWA research record},
  url = {https://whitzard.tech/nuwa/whitzard-index/work/ape/},
  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--ape.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--ape.json.}
}

证据快照 2026-10-29.1 · 索引更新 · RSS

参考文献

[1] Matthew Kowal, Jasper Timm 等. It's the Thought that Counts: Evaluating the Attempts of Frontier LLMs to Persuade on Harmful Topics. 2026-02-15. v4 [arXiv]; Author affiliations on the original paper title page; Abstract; §3 Methodology