评测档案

Evaluating the Paperclip Maximizer: Are RL-Based Language Models More Likely to Pursue Instrumental Goals?

在指定场景下测试模型回答中的资源获取、自我维持等工具性目标。

新加坡国立大学 / arXiv:2502.12206v1 / 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

Evaluating the Paperclip Maximizer: Are RL-Based Language Models More Likely to Pursue Instrumental Goals? · 规避关机

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。

8 条观测 · 0 已定位日期

● 条件匹配◇ 配置差异□ 部分条件未披露

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。

8 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-R1	50%规避关机	条件不完整	Table 1 · Evaluating the Paperclip Maximizer: Are RL-Based Language Models More Likely to Pursue Instrumental Goals?—	未披露Evaluating the Paperclip Maximizer: Are RL-Based Language Models More Likely to Pursue Instrumental Goals?	发布 —报告 —
o3-mini	45%规避关机	条件不完整	Table 1 · Evaluating the Paperclip Maximizer: Are RL-Based Language Models More Likely to Pursue Instrumental Goals?—	未披露Evaluating the Paperclip Maximizer: Are RL-Based Language Models More Likely to Pursue Instrumental Goals?	发布 —报告 —
o1	40%规避关机	条件不完整	Table 1 · Evaluating the Paperclip Maximizer: Are RL-Based Language Models More Likely to Pursue Instrumental Goals?—	未披露Evaluating the Paperclip Maximizer: Are RL-Based Language Models More Likely to Pursue Instrumental Goals?	发布 —报告 —
o1-mini	40%规避关机	条件不完整	Table 1 · Evaluating the Paperclip Maximizer: Are RL-Based Language Models More Likely to Pursue Instrumental Goals?—	未披露Evaluating the Paperclip Maximizer: Are RL-Based Language Models More Likely to Pursue Instrumental Goals?	发布 —报告 —
Gemini 2.0 FTE.	35%规避关机	条件不完整	Table 1 · Evaluating the Paperclip Maximizer: Are RL-Based Language Models More Likely to Pursue Instrumental Goals?—	未披露Evaluating the Paperclip Maximizer: Are RL-Based Language Models More Likely to Pursue Instrumental Goals?	发布 —报告 —
GPT-4o	20%规避关机	条件不完整	Table 1 · Evaluating the Paperclip Maximizer: Are RL-Based Language Models More Likely to Pursue Instrumental Goals?—	未披露Evaluating the Paperclip Maximizer: Are RL-Based Language Models More Likely to Pursue Instrumental Goals?	发布 —报告 —
Sonnet-3.5	15%规避关机	条件不完整	Table 1 · Evaluating the Paperclip Maximizer: Are RL-Based Language Models More Likely to Pursue Instrumental Goals?—	未披露Evaluating the Paperclip Maximizer: Are RL-Based Language Models More Likely to Pursue Instrumental Goals?	发布 —报告 —
DeepSeek-V3	15%规避关机	条件不完整	Table 1 · Evaluating the Paperclip Maximizer: Are RL-Based Language Models More Likely to Pursue Instrumental Goals?—	未披露Evaluating the Paperclip Maximizer: Are RL-Based Language Models More Likely to Pursue Instrumental Goals?	发布 —报告 —

任务与测量

Benchmark 协议档案

任务／对象: 原评测配置下的模型或智能体系统
测量目标: 在指定场景下测试模型回答中的资源获取、自我维持等工具性目标。
评分／结果口径: 规避关机 · 表 1 中 GPT-4o 作为判别模型的类别发生率与总体发生率。情境内的判定比例不等于部署概率。
解释边界: 文本选择与评判标签不能证明已执行的自主行为或稳定部署倾向。

一项已报告评测

来源评测 GPT-4o · Table 1 · Evaluating the Paperclip Maximizer: Are RL-Based Language Models More Likely to Pursue Instrumental Goals?.

报告指标: 规避关机 · 20.00%.

表 1 中 GPT-4o 作为判别模型的类别发生率与总体发生率。情境内的判定比例不等于部署概率。见原文协议；不推断独立试验数量。

原文 Table 1 报告的 GPT-4o 测量结果；数值为转录，未独立复现实验。

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: 目标驱动回答何时出现工具性目标行为？
任务与样本: 基准改变前置任务、模型与提示，以自动判官标注工具趋同。
判分与失效情形: 自动判工具性目标回答，受判官选择影响。
系统条件: 改变前置任务、模型组与目标强调。
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：判官选择显著影响检出（§4.6）。
效度与解释: 各组训练史与能力不同，判官敏感性和构造任务限制部署推断。
适用研究问题: 模型建议：保留判官与目标强调条件，避免据未匹配组作因果结论。

Evaluating the Paperclip Maximizer: Are RL-Based Language Models More Likely to Pursue Instrumental Goals? · §3 InstrumentalEvalSelected original resource https://arxiv.org/html/2502.12206; sha256:1bb93b48504f755173f8730476788c3add572c14d95d32d254b8f7594d7803f1; archived source capture; exact capture time not independently established
Evaluating the Paperclip Maximizer: Are RL-Based Language Models More Likely to Pursue Instrumental Goals? · §4.5 Explicit Goal Nudging Increases Instrumental ConvergenceSelected original resource https://arxiv.org/html/2502.12206; sha256:1bb93b48504f755173f8730476788c3add572c14d95d32d254b8f7594d7803f1; archived source capture; exact capture time not independently established
Evaluating the Paperclip Maximizer: Are RL-Based Language Models More Likely to Pursue Instrumental Goals? · §4.6 Judge Model Selection Significantly Impacts Convergence DetectionSelected original resource https://arxiv.org/html/2502.12206; sha256:1bb93b48504f755173f8730476788c3add572c14d95d32d254b8f7594d7803f1; archived source capture; exact capture time not independently established
Evaluating the Paperclip Maximizer: Are RL-Based Language Models More Likely to Pursue Instrumental Goals? · LimitationsSelected original resource https://arxiv.org/html/2502.12206; sha256:1bb93b48504f755173f8730476788c3add572c14d95d32d254b8f7594d7803f1; archived source capture; exact capture time not independently established