白泽（Whitzard）女娲实验室 ↗
NUWANUWA INDEX

搜索 NUWA Index

研究 · 测量协议 · 评测结果 · 观点

English
← NUWA Index

评测档案

Attempt to Persuade Eval (APE)

测量模型在模拟对话中针对特定有害议题的说服尝试。

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 / arXiv:2506.02873v4 / Propensity · 行为倾向

来源与出处 ↓ 阅读研究记录 →

跨报告监测

已发表证据

Attempt to Persuade Eval (APE) · Turn-1 persuasion attempt 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。

32 条观测 · 24 已定位日期

● 条件匹配◇ 配置差异□ 部分条件未披露
Turn-1 persuasion attempt rate (%) DeepSeek V4 Flash Claude Sonnet 5 Kimi K3 GPT-5.6 Terra 0 20 40 60 80 100 94% 33% 11% 7% 阴谋论

8 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。

32 条观测

测量值与来源条件
系统／模型结果条件状态任务／条件评测方／报告日期
DeepSeek V4 FlashDeepSeek V4 Flash 073194%Turn-1 persuasion attempt rate条件不完整APE topic categories阴谋论600 topics,3 persuasion turns each—Neo ResearchEvaluating Kimi K3 for Misalignment Risks发布 —报告 2026-09-11
Claude Sonnet 533%Turn-1 persuasion attempt rate条件不完整APE topic categories阴谋论600 topics,3 persuasion turns each—Neo ResearchEvaluating Kimi K3 for Misalignment Risks发布 2026-06-30报告 2026-09-11
Kimi K311%Turn-1 persuasion attempt rate条件不完整APE topic categories阴谋论600 topics,3 persuasion turns each—Neo ResearchEvaluating Kimi K3 for Misalignment Risks发布 2026-07-16报告 2026-09-11
GPT-5.6 Terra7%Turn-1 persuasion attempt rate条件不完整APE topic categories阴谋论600 topics,3 persuasion turns each—Neo ResearchEvaluating Kimi K3 for Misalignment Risks发布 2026-07-09报告 2026-09-11
DeepSeek V4 FlashDeepSeek V4 Flash 073199%Turn-1 persuasion attempt rate条件不完整APE topic categories削弱控制600 topics,3 persuasion turns each—Neo ResearchEvaluating Kimi K3 for Misalignment Risks发布 —报告 2026-09-11
Claude Sonnet 532%Turn-1 persuasion attempt rate条件不完整APE topic categories削弱控制600 topics,3 persuasion turns each—Neo ResearchEvaluating Kimi K3 for Misalignment Risks发布 2026-06-30报告 2026-09-11
GPT-5.6 Terra28%Turn-1 persuasion attempt rate条件不完整APE topic categories削弱控制600 topics,3 persuasion turns each—Neo ResearchEvaluating Kimi K3 for Misalignment Risks发布 2026-07-09报告 2026-09-11
Kimi K35%Turn-1 persuasion attempt rate条件不完整APE topic categories削弱控制600 topics,3 persuasion turns each—Neo ResearchEvaluating Kimi K3 for Misalignment Risks发布 2026-07-16报告 2026-09-11
DeepSeek V4 FlashDeepSeek V4 Flash 073157%Turn-1 persuasion attempt rate条件不完整APE topic categories明确有害话题（总体）600 topics,3 persuasion turns each—Neo ResearchEvaluating Kimi K3 for Misalignment Risks发布 —报告 2026-09-11
Claude Sonnet 52%Turn-1 persuasion attempt rate条件不完整APE topic categories明确有害话题（总体）600 topics,3 persuasion turns each—Neo ResearchEvaluating Kimi K3 for Misalignment Risks发布 2026-06-30报告 2026-09-11
Kimi K30%Turn-1 persuasion attempt rate条件不完整APE topic categories明确有害话题（总体）600 topics,3 persuasion turns each—Neo ResearchEvaluating Kimi K3 for Misalignment Risks发布 2026-07-16报告 2026-09-11
GPT-5.6 Terra0%Turn-1 persuasion attempt rate条件不完整APE topic categories明确有害话题（总体）600 topics,3 persuasion turns each—Neo ResearchEvaluating Kimi K3 for Misalignment Risks发布 2026-07-09报告 2026-09-11
DeepSeek V4 FlashDeepSeek V4 Flash 073170%Turn-1 persuasion attempt rate条件不完整APE topic categories实施暴力600 topics,3 persuasion turns each—Neo ResearchEvaluating Kimi K3 for Misalignment Risks发布 —报告 2026-09-11
Claude Sonnet 55%Turn-1 persuasion attempt rate条件不完整APE topic categories实施暴力600 topics,3 persuasion turns each—Neo ResearchEvaluating Kimi K3 for Misalignment Risks发布 2026-06-30报告 2026-09-11
Kimi K30%Turn-1 persuasion attempt rate条件不完整APE topic categories实施暴力600 topics,3 persuasion turns each—Neo ResearchEvaluating Kimi K3 for Misalignment Risks发布 2026-07-16报告 2026-09-11
GPT-5.6 Terra0%Turn-1 persuasion attempt rate条件不完整APE topic categories实施暴力600 topics,3 persuasion turns each—Neo ResearchEvaluating Kimi K3 for Misalignment Risks发布 2026-07-09报告 2026-09-11
DeepSeek V4 FlashDeepSeek V4 Flash 073155%Turn-1 persuasion attempt rate条件不完整APE topic categories人口贩运600 topics,3 persuasion turns each—Neo ResearchEvaluating Kimi K3 for Misalignment Risks发布 —报告 2026-09-11
Kimi K30%Turn-1 persuasion attempt rate条件不完整APE topic categories人口贩运600 topics,3 persuasion turns each—Neo ResearchEvaluating Kimi K3 for Misalignment Risks发布 2026-07-16报告 2026-09-11
GPT-5.6 Terra0%Turn-1 persuasion attempt rate条件不完整APE topic categories人口贩运600 topics,3 persuasion turns each—Neo ResearchEvaluating Kimi K3 for Misalignment Risks发布 2026-07-09报告 2026-09-11
Claude Sonnet 50%Turn-1 persuasion attempt rate条件不完整APE topic categories人口贩运600 topics,3 persuasion turns each—Neo ResearchEvaluating Kimi K3 for Misalignment Risks发布 2026-06-30报告 2026-09-11
DeepSeek V4 FlashDeepSeek V4 Flash 073150%Turn-1 persuasion attempt rate条件不完整APE topic categories大规模杀戮600 topics,3 persuasion turns each—Neo ResearchEvaluating Kimi K3 for Misalignment Risks发布 —报告 2026-09-11
Kimi K30%Turn-1 persuasion attempt rate条件不完整APE topic categories大规模杀戮600 topics,3 persuasion turns each—Neo ResearchEvaluating Kimi K3 for Misalignment Risks发布 2026-07-16报告 2026-09-11
GPT-5.6 Terra0%Turn-1 persuasion attempt rate条件不完整APE topic categories大规模杀戮600 topics,3 persuasion turns each—Neo ResearchEvaluating Kimi K3 for Misalignment Risks发布 2026-07-09报告 2026-09-11
Claude Sonnet 50%Turn-1 persuasion attempt rate条件不完整APE topic categories大规模杀戮600 topics,3 persuasion turns each—Neo ResearchEvaluating Kimi K3 for Misalignment Risks发布 2026-06-30报告 2026-09-11
DeepSeek V4 FlashDeepSeek V4 Flash 073135%Turn-1 persuasion attempt rate条件不完整APE topic categories针对未成年人的性侵害600 topics,3 persuasion turns each—Neo ResearchEvaluating Kimi K3 for Misalignment Risks发布 —报告 2026-09-11
Kimi K30%Turn-1 persuasion attempt rate条件不完整APE topic categories针对未成年人的性侵害600 topics,3 persuasion turns each—Neo ResearchEvaluating Kimi K3 for Misalignment Risks发布 2026-07-16报告 2026-09-11
GPT-5.6 Terra0%Turn-1 persuasion attempt rate条件不完整APE topic categories针对未成年人的性侵害600 topics,3 persuasion turns each—Neo ResearchEvaluating Kimi K3 for Misalignment Risks发布 2026-07-09报告 2026-09-11
Claude Sonnet 50%Turn-1 persuasion attempt rate条件不完整APE topic categories针对未成年人的性侵害600 topics,3 persuasion turns each—Neo ResearchEvaluating Kimi K3 for Misalignment Risks发布 2026-06-30报告 2026-09-11
DeepSeek V4 FlashDeepSeek V4 Flash 073175%Turn-1 persuasion attempt rate条件不完整APE topic categories虐待无辜者600 topics,3 persuasion turns each—Neo ResearchEvaluating Kimi K3 for Misalignment Risks发布 —报告 2026-09-11
Claude Sonnet 55%Turn-1 persuasion attempt rate条件不完整APE topic categories虐待无辜者600 topics,3 persuasion turns each—Neo ResearchEvaluating Kimi K3 for Misalignment Risks发布 2026-06-30报告 2026-09-11
Kimi K30%Turn-1 persuasion attempt rate条件不完整APE topic categories虐待无辜者600 topics,3 persuasion turns each—Neo ResearchEvaluating Kimi K3 for Misalignment Risks发布 2026-07-16报告 2026-09-11
GPT-5.6 Terra0%Turn-1 persuasion attempt rate条件不完整APE topic categories虐待无辜者600 topics,3 persuasion turns each—Neo ResearchEvaluating Kimi K3 for Misalignment Risks发布 2026-07-09报告 2026-09-11

实际评测方: Neo Research
报告发布方: Neo Research

任务与测量

Benchmark 协议档案

任务／对象
模拟多轮说服对话中的模型
测量目标
特定对话条件下说服尝试的发生率。
评分／结果口径
Turn-1 persuasion attempt rate · 记录配置下的来源报告指标。
解释边界
模拟尝试不等于已测量的人类信念变化或大规模虚假信息传播。

一项已报告评测

来源评测 Kimi K3 · APE topic categories.

运行配置: 600 topics,3 persuasion turns each.

报告指标: Turn-1 persuasion attempt rate · 11%.

记录配置下的来源报告指标。 原文未披露本结果的样本量

Turn-1 persuasion attempt rate：11 %；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
未观测说服成功时，模型是否仍尝试有害说服？
任务与样本
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。
判分与失效情形
说服尝试频率，不是模拟或人类说服成功。
系统条件
多轮模型对话及自动尝试判分。
统计含义
来源报告的结果表述；这不等同于已确定运行次数、分母或估计方法： 目标是说服意愿，不是信念变化（§2.2）。
效度与解释
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）。
适用研究问题
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。

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

32 条结果记录 · 4 个报告系统名称

协议与比较依据 (8)
Attempt to Persuade Eval (APE)：Turn-1 persuasion attempt rate · Neo Researchsource-report · 范围已审查
Attempt to Persuade Eval (APE)：Turn-1 persuasion attempt rate · Neo Researchsource-report · 范围已审查
Attempt to Persuade Eval (APE)：Turn-1 persuasion attempt rate · Neo Researchsource-report · 范围已审查
Attempt to Persuade Eval (APE)：Turn-1 persuasion attempt rate · Neo Researchsource-report · 范围已审查
Attempt to Persuade Eval (APE)：Turn-1 persuasion attempt rate · Neo Researchsource-report · 范围已审查
Attempt to Persuade Eval (APE)：Turn-1 persuasion attempt rate · Neo Researchsource-report · 范围已审查
Attempt to Persuade Eval (APE)：Turn-1 persuasion attempt rate · Neo Researchsource-report · 范围已审查
Attempt to Persuade Eval (APE)：Turn-1 persuasion attempt rate · Neo Researchsource-report · 范围已审查

Neo Research

Evaluating DeepSeek v4 Pro for Frontier Risks →

实际评测方: Neo Research

19 条结果记录 · 4 个报告系统名称

协议与比较依据 (16)
Attempt to Persuade Eval (APE)：attempt rate · Neo Researchsource-report · 范围已审查
Attempt to Persuade Eval (APE)：attempt rate · Neo Researchsource-report · 范围已审查
Attempt to Persuade Eval (APE)：attempt rate · Neo Researchsource-report · 范围已审查
Attempt to Persuade Eval (APE)：attempt rate · Neo Researchsource-report · 范围已审查
Attempt to Persuade Eval (APE)：attempt rate · Neo Researchsource-report · 范围已审查
Attempt to Persuade Eval (APE)：no attempt count · Neo Researchsource-report · 范围已审查
Attempt to Persuade Eval (APE)：no attempt count · Neo Researchsource-report · 范围已审查
Attempt to Persuade Eval (APE)：no attempt count · Neo Researchsource-report · 范围已审查
Attempt to Persuade Eval (APE)：no attempt count · Neo Researchsource-report · 范围已审查
Attempt to Persuade Eval (APE)：no attempt count · Neo Researchsource-report · 范围已审查
Attempt to Persuade Eval (APE)：NoncontroversiallyHarmful acceptance rate · Neo Researchsource-report · 范围已审查
Attempt to Persuade Eval (APE)：refusal rate · Neo Researchsource-report · 范围已审查
Attempt to Persuade Eval (APE)：refusal rate · Neo Researchsource-report · 范围已审查
Attempt to Persuade Eval (APE)：refusal rate · Neo Researchsource-report · 范围已审查
Attempt to Persuade Eval (APE)：refusal rate · Neo Researchsource-report · 范围已审查
Attempt to Persuade Eval (APE)：refusal rate · Neo Researchsource-report · 范围已审查

SaferAI

GLM-5.2 Risk Evaluation Report →

实际评测方: SaferAI

6 条结果记录 · 3 个报告系统名称

协议与比较依据 (6)
Attempt to Persuade Eval (APE)：APE Turn-1 persuasion-attempt rate · SaferAIsource-report · 范围已审查
Attempt to Persuade Eval (APE)：APE Turn-1 persuasion-attempt rate · SaferAIsource-report · 范围已审查
Attempt to Persuade Eval (APE)：APE Turn-1 persuasion-attempt rate · SaferAIsource-report · 范围已审查
Attempt to Persuade Eval (APE)：APE Turn-1 persuasion-attempt rate · SaferAIsource-report · 范围已审查
Attempt to Persuade Eval (APE)：APE Turn-1 persuasion-attempt rate · SaferAIsource-report · 范围已审查
Attempt to Persuade Eval (APE)：APE Turn-1 persuasion-attempt rate · SaferAIsource-report · 范围已审查
原始来源图表与材料
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。 查看原图 ↗

评估出处

Evaluating Kimi K3 for Misalignment Risks → · 外部报告

Evaluating DeepSeek v4 Pro for Frontier Risks → · 外部报告

Early Results on Evaluating GLM 5.2 for Behavioral and Persuasion Risks → · 外部报告

GLM-5.2 Risk Evaluation Report → · 外部报告

相关研究问题

参考文献

  1. [1] Kimi Code — What’s New ; Kimi K3 · July 16, 2026
  2. [2] kimi-k3 developer model card ; Model Summary: Total Parameters; Activated Parameters
  3. [3] Z.ai release notes ; GLM-5.2 entry dated 2026-06-16
  4. [4] GLM-5.2 model card ; Parameter counts not disclosed in reviewed developer source
  5. [5] deepseek-v4-pro release ; News article title, dated heading and availability
  6. [6] deepseek-v4-pro developer model card ; Introduction; model comparison table
  7. [7] GPT-5.6 official general availability ; July 9, 2026: general availability following limited trusted-partner preview; Parameter counts not disclosed in reviewed launch article
  8. [8] Introducing Claude Sonnet 5 — Anthropic ; Publication date and availability section; Official launch article; parameter counts not disclosed
  9. [9] OpenAI — introducing gpt 5 5 ; Publication date and availability section; Parameter counts not disclosed in launch article
  10. [10] Introducing Claude Opus 4.7 — Anthropic ; Publication date and availability section; Official launch article; parameter counts not disclosed
  11. [11] Introducing Sonnet 4.6 — Anthropic ; Publication date and availability section; Official launch article; parameter counts not disclosed
  12. [12] Gemini 2.5: Our newest Gemini model with thinking ; Publication date and availability section; Official launch article; parameter counts not disclosed
  13. [13] OpenAI — hello gpt 4o ; Publication date and availability section; Parameter counts not disclosed in launch article
  14. [14] Evaluating Kimi K3 for Misalignment Risks. 2026-09-11. September 2026 ; §7.1 Table 16 p.58
  15. [15] GLM-5.2 Risk Evaluation Report. 2026-08-02. v1.1 · revised 26 August 2026 ; p.22 §5.2/Fig.10; p.43 Table A.4 category size
  16. [16] Early Results on Evaluating GLM 5.2 for Behavioral and Persuasion Risks. 2026-07-03. 3 July 2026 ; #persuasion-benchmarks; Figure 5 caption and following paragraph
  17. [17] Evaluating DeepSeek v4 Pro for Frontier Risks. 2026-05-29. v1.1 · 29 May 2026 ; PDF p. 16; APE noncontroversially harmful: Physical violence; refusal rate; PDF p. 16; APE noncontroversially harmful: Physical violence; no attempt count
  18. [18] 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; Figure 3