NUWA 评测档案
PersuSafety
模型参与说服请求的方式如何随主题危害变化？
跨报告监测
已发表证据
PersuSafety · 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。
12 条观测 · 9 已定位日期
● 条件匹配◇ 配置差异□ 部分条件未披露
3 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。
12 条观测
| 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 0731 | 100%Turn-1 persuasion attempt rate | 条件不完整 | PersuSafety open-ended single-turn adaptation低危害— | Neo ResearchEvaluating Kimi K3 for Misalignment Risks | 发布 —报告 2026-09-11 |
 Claude Sonnet 5 | 89%Turn-1 persuasion attempt rate | 条件不完整 | PersuSafety open-ended single-turn adaptation低危害— | Neo ResearchEvaluating Kimi K3 for Misalignment Risks | 发布 2026-06-30报告 2026-09-11 |
 GPT-5.6 Terra | 86%Turn-1 persuasion attempt rate | 条件不完整 | PersuSafety open-ended single-turn adaptation低危害— | Neo ResearchEvaluating Kimi K3 for Misalignment Risks | 发布 2026-07-09报告 2026-09-11 |
 Kimi K3 | 75%Turn-1 persuasion attempt rate | 条件不完整 | PersuSafety open-ended single-turn adaptation低危害— | Neo ResearchEvaluating Kimi K3 for Misalignment Risks | 发布 2026-07-16报告 2026-09-11 |
| DeepSeek V4 FlashDeepSeek V4 Flash 0731 | 84%Turn-1 persuasion attempt rate | 条件不完整 | PersuSafety open-ended single-turn adaptation中等危害— | Neo ResearchEvaluating Kimi K3 for Misalignment Risks | 发布 —报告 2026-09-11 |
| Kimi K3 | 30%Turn-1 persuasion attempt rate | 条件不完整 | PersuSafety open-ended single-turn adaptation中等危害— | Neo ResearchEvaluating Kimi K3 for Misalignment Risks | 发布 2026-07-16报告 2026-09-11 |
| Claude Sonnet 5 | 19%Turn-1 persuasion attempt rate | 条件不完整 | PersuSafety open-ended single-turn adaptation中等危害— | Neo ResearchEvaluating Kimi K3 for Misalignment Risks | 发布 2026-06-30报告 2026-09-11 |
| GPT-5.6 Terra | 5%Turn-1 persuasion attempt rate | 条件不完整 | PersuSafety open-ended single-turn adaptation中等危害— | Neo ResearchEvaluating Kimi K3 for Misalignment Risks | 发布 2026-07-09报告 2026-09-11 |
| DeepSeek V4 FlashDeepSeek V4 Flash 0731 | 86%Turn-1 persuasion attempt rate | 条件不完整 | PersuSafety open-ended single-turn adaptation高危害— | Neo ResearchEvaluating Kimi K3 for Misalignment Risks | 发布 —报告 2026-09-11 |
| Kimi K3 | 31%Turn-1 persuasion attempt rate | 条件不完整 | PersuSafety open-ended single-turn adaptation高危害— | Neo ResearchEvaluating Kimi K3 for Misalignment Risks | 发布 2026-07-16报告 2026-09-11 |
| Claude Sonnet 5 | 17%Turn-1 persuasion attempt rate | 条件不完整 | PersuSafety open-ended single-turn adaptation高危害— | Neo ResearchEvaluating Kimi K3 for Misalignment Risks | 发布 2026-06-30报告 2026-09-11 |
| GPT-5.6 Terra | 0%Turn-1 persuasion attempt rate | 条件不完整 | PersuSafety open-ended single-turn adaptation高危害— | Neo ResearchEvaluating Kimi K3 for Misalignment Risks | 发布 2026-07-09报告 2026-09-11 |
结合任务范围与运行配置解读结果。
METR Time Horizon 1.1 来源已停止积极更新，最后更新于 2026 年 9 月 8 日。超过 16 小时的估计超出其可靠测量范围。
协议、配置与原始证据 →任务与测量
Benchmark 协议档案
- 任务／对象
- 给定任务配置中的模型或智能体。
- 测量目标
- 定义说服情境、模拟对话与安全评测；改造判别器作为独立变体。
- 评分／结果口径
- Medium-harm attempt rate · 定义说服情境、模拟对话与安全评测；改造判别器作为独立变体。
- 解释边界
- 参与意愿或模拟对话评分不直接测量人的信念变化。
定义说服情境、模拟对话与安全评测；改造判别器作为独立变体。
一项已报告评测
来源评测 GLM-5.2 · PersuSafety adapted free-generation judge.
报告指标: Medium-harm attempt rate · 54%.
定义说服情境、模拟对话与安全评测；改造判别器作为独立变体。 见报告特定样本与分母。
原文 §2; Figure 6 报告的测量；转录不等于独立复现实验。
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
- 模型会拒绝不伦理说服，并在接任务后避免不伦理策略吗？
- 任务与样本
- 三阶段框架生成任务、模拟对话并判分，涵盖六话题十五策略；八说服者默认以GPT-4o模拟受众。
- 判分与失效情形
- 拒绝及不伦理策略分，策略分析以未拒任务为条件。
- 系统条件
- 八说服模型，默认GPT-4o模拟受众及脆弱特征。
- 统计含义
- 现有材料尚未确定
- 效度与解释
- 模拟受众不确立人类影响，条件策略分不能当无条件发生率。
- 适用研究问题
- 模型建议：并列报告拒绝与条件策略使用，并保留受众特征可见性。
- PersuSafety · §3.3 Benchmarked LLMs and SetupSelected original resource https://arxiv.org/html/2504.10430; sha256:b97a60a11e25b2c2f1283620ee8ca36bb2c583e7fba5f06d6982564475a973f4; archived source capture; exact capture time not independently established
- PersuSafety · §4 Experiments and Discussions: denominator noteSelected original resource https://arxiv.org/html/2504.10430; sha256:b97a60a11e25b2c2f1283620ee8ca36bb2c583e7fba5f06d6982564475a973f4; archived source capture; exact capture time not independently established
- PersuSafety · §4.2 How LLM Persuaders Exploit Persuadee’s Vulnerabilities?Selected original resource https://arxiv.org/html/2504.10430; sha256:b97a60a11e25b2c2f1283620ee8ca36bb2c583e7fba5f06d6982564475a973f4; archived source capture; exact capture time not independently established
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
Early Results on Evaluating GLM 5.2 for Behavioral and Persuasion Risks →
实际评测方: Neo Research
9 条结果记录 · 4 个报告系统名称
协议与比较依据 (6)
PersuSafety adapted free-generation judge · High-harm attempt ratesource-report · 范围已审查
PersuSafety adapted free-generation judge · Medium-harm attempt ratesource-report · 范围已审查
PersuSafety：High-harm attempt rate · Neo Researchsource-report · 范围已审查
PersuSafety：Medium-harm attempt rate · Neo Researchsource-report · 范围已审查
PersuSafety：Modified-PersuSafety persuasion-attempt rate · Neo Researchsource-report · 范围已审查
PersuSafety：Modified-PersuSafety persuasion-attempt rate · Neo Researchsource-report · 范围已审查
Neo Research
Evaluating Kimi K3 for Misalignment Risks →
实际评测方: Neo Research
12 条结果记录 · 4 个报告系统名称
协议与比较依据 (3)
PersuSafety：Turn-1 persuasion attempt rate · Neo Researchsource-report · 范围已审查
PersuSafety：Turn-1 persuasion attempt rate · Neo Researchsource-report · 范围已审查
PersuSafety：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 · 未披露
17 条结果记录 · 6 个报告系统名称
协议与比较依据 (12)
PersuSafety：acceptance rate · Neo Researchsource-report · 范围已审查
PersuSafety：acceptance rate · Neo Researchsource-report · 范围已审查
PersuSafety：accepted high-harm count · Neo Researchsource-report · 范围已审查
PersuSafety：accepted low-harm count · Neo Researchsource-report · 范围已审查
PersuSafety：accepted medium-harm count · Neo Researchsource-report · 范围已审查
PersuSafety：accepted task count · Neo Researchsource-report · 范围已审查
PersuSafety：high-harm attempt rate · Neo Researchsource-report · 范围已审查
PersuSafety：low-harm attempt rate · Neo Researchsource-report · 范围已审查
PersuSafety：medium-harm attempt rate · Neo Researchsource-report · 范围已审查
PersuSafety：overall attempt rate · Neo Researchsource-report · 范围已审查
PersuSafety：refusal rate · Neo Researchsource-report · 范围已审查
PersuSafety：refusal rate · Neo Researchsource-report · 范围已审查
原始来源图表与材料
PersuSafety adapted free-generation judge
原始评测方结果 · 来源顺序
筛选系统
Medium-harm attempt rate
0–100%
GLM-5.2GLM-5.2 · 54% · 定义说服情境、模拟对话与安全评测；改造判别器作为独立变体。 · 见报告特定样本与分母。
54%Claude Sonnet 5Claude Sonnet 5 · 57% · 定义说服情境、模拟对话与安全评测；改造判别器作为独立变体。 · 见报告特定样本与分母。
57%DeepSeek V4 ProDeepSeek V4 Pro · 76% · 定义说服情境、模拟对话与安全评测；改造判别器作为独立变体。 · 见报告特定样本与分母。
76%显示 3 / 3 个系统
定义说服情境、模拟对话与安全评测；改造判别器作为独立变体。
参与意愿或模拟对话评分不直接测量人的信念变化。 本评估显示报告日期和原始执行方。
评测条件与系统配置
- GLM-5.2
- 见报告特定样本与分母。 · PersuSafety adapted free-generation judge
- modelVersion: 未注明 · harness: 未注明 · tools: 未注明 · budget: 未注明
- Claude Sonnet 5
- 见报告特定样本与分母。 · PersuSafety adapted free-generation judge
- modelVersion: 未注明 · harness: 未注明 · tools: 未注明 · budget: 未注明
- DeepSeek V4 Pro
- 见报告特定样本与分母。 · PersuSafety adapted free-generation judge
- modelVersion: 未注明 · harness: 未注明 · tools: 未注明 · budget: 未注明
| 系统 / 配置 | Medium-harm attempt rate | High-harm attempt rate | 发表 / 核验日期 |
|---|---|---|---|
| GLM-5.2未注明 Harness | 54% | 6% | 发表2026-07-03 |
| Claude Sonnet 5未注明 Harness | 57% | 31% | 发表2026-07-03 |
| DeepSeek V4 Pro未注明 Harness | 76% | 78% | 发表2026-07-03 |
评估出处
Early Results on Evaluating GLM 5.2 for Behavioral and Persuasion Risks → · 外部报告
相关研究问题
虚假信息与影响 / 说服尝试 →
该协议明确测量该问题或其使能子能力，结论保留给定条件。
参考文献
- [1] Kimi Code — What’s New ; Kimi K3 · July 16, 2026
- [2] kimi-k3 developer model card ; Model Summary: Total Parameters; Activated Parameters
- [3] Z.ai release notes ; GLM-5.2 entry dated 2026-06-16
- [4] GLM-5.2 model card ; Parameter counts not disclosed in reviewed developer source
- [5] deepseek-v4-pro release ; News article title, dated heading and availability
- [6] deepseek-v4-pro developer model card ; Introduction; model comparison table
- [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] Introducing Claude Sonnet 5 — Anthropic ; Publication date and availability section; Official launch article; parameter counts not disclosed
- [9] Gemini 2.5: Our newest Gemini model with thinking ; Publication date and availability section; Official launch article; parameter counts not disclosed
- [10] Evaluating Kimi K3 for Misalignment Risks. 2026-09-11. September 2026 ; §7.2 Figure 40 p.60
- [11] Early Results on Evaluating GLM 5.2 for Behavioral and Persuasion Risks. 2026-07-03. 3 July 2026 ; §2; Figure 6; #persuasion-benchmarks; Figure 6 caption and following two paragraphs
- [12] Evaluating DeepSeek v4 Pro for Frontier Risks. 2026-05-29. v1.1 · 29 May 2026 ; PDF p. 14; PersuSafety single turn; refusal rate; PDF p. 15; PersuSafety single turn; acceptance rate; PDF p. 15; PersuSafety single turn; accepted low-harm count
- [13] PersuSafety. 2025-04-14