前沿模型偏爱数字 7：Astra 等 30/30 次选 7

An Vo @an_vo12原推发布于 2026.10.16

在 X 上打开原推

The bias seems even stronger in frontier models! GPT-6 Astra, Claude Opus 5.5, and Gemini 4 Argon all picked 7 in 30/30 times 🤯 But WHY exactly do LLMs love the number 7? 👀 It might have something to do with **SFT training data**! In our #EMNLP2026 paper (https://deepbias.github.io), we found that Olmo 3 7B picked a mix of numbers before SFT, BUT after SFT it picked 7 in 26/30 times. The SFT training data is full of examples where the answer is 7 👇

这条是引用着另一条写的：x.com/adamzwasserman/status/2110135217721880935