原帖引用 · AI 翻译
标题：“大语言模型中随机性的幻觉”
Google DeepMind 的一篇新论文发现，AI 模型无法自行做出真正随机的选择，但如果计算机程序先给较大的模型一个随机数，它们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，例如以不可预测的顺序检查服务器。大语言模型声称会均匀随机地选择，却总是偏向 7、42 或答案 C 等选项。
研究人员要求 Gemini、Qwen3 和 OLMo 模型分别从简单分布中采样 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，让它们将其转换为所需的选择。
报告中的每项检验是否符合目标概率的统计测试均未通过，改变温度或 top-p 也未能解决问题。在一场安全博弈中，真正随机的检查会让攻击者有 75% 的获胜概率，而 Gemini 的选择让这一概率达到 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将其传入，而不要让模型自己随机选择。
原文
New Google DeepMind paper finds that AI models cannot make truly random choices on their own, but larger models pick correctly when a computer program gives them a random number first. Agents often need to pick actions at random with set odds, like checking servers in an unpredictable order. LLMs say they will pick uniformly, then keep landing on favorites like 7, 42, or answer C. They asked Gemini, Qwen3, and OLMo models to sample from simple distributions 1024 times each. They also tried other temperature and sampling settings, and gave models a random number to convert. Every reported statistical test for matching the target odds failed, and changing temperature or top-p did not fix it. In a security game where truly random checks let attackers win 75% of the time, Gemini's picks let them win 96.6% to 99.9%. If your agent needs randomness, generate it in code and pass it in, instead of asking the model to be random.