aifollow.news Search
Back Rohan Paul
Rohan Paul· @rohanpaul_ai · X· · Original publication time AI score54

Google DeepMind paper finds AI models favor certain answers when asked to choose randomlyMachine translation

Automatically verified and published · Generated and evidence-checked automatically; not reviewed by a human.

AI introduction

Researchers tested Gemini, Qwen3, and OLMo models with 1,024 samples each. Every reported statistical test failed to show that their choices matched the target odds, and changing temperature or top-p did not fix the problem. In a security game, the models' choices let attackers win 96.6% to 99.9% of the time, compared with 75% for truly random checks. Larger models picked correctly when first given a random number generated by a computer program.

Original post · AI translation

@rohanpaul_ai
Translation · refer to the original

Google DeepMind 的一篇新论文发现,AI 模型无法自行作出真正随机的选择,但如果计算机程序先给出一个随机数,较大的模型就能按要求正确选择。

智能体经常需要按设定概率随机选择行动,比如以不可预测的顺序检查服务器。大语言模型会说自己将均匀随机选择,却总是选中 7、42 或答案 C 之类的偏好选项。

研究人员让 Gemini、Qwen3 和 OLMo 模型分别从简单分布中抽样 1024 次。他们还尝试了其他温度和采样设置,并给模型一个随机数让其转换。

所有报告的检验结果都未能证明抽样符合目标概率,改变温度或 top-p 也未能解决问题。在一项安全博弈中,如果检查真正随机,攻击者的胜率为 75%;而使用 Gemini 的选择时,攻击者的胜率达到 96.6% 至 99.9%。

如果你的智能体需要随机性,就用代码生成随机数并传给它,而不是要求模型自行作出随机选择。

Source: Rohan Paul · x.com

Research
Found an error? Send a correction