告诉 HN：Gemini 3.5 Flash 存在一些愚蠢的故障
6 分•作者: XCSme•大约 2 个月前
我以为自己要疯了，想用 Gemini 3.5 Flash 给一些答案打分，但它总是给正确答案打 7 分，而不是 10 分。<p>显然，一旦你添加了“评分标准”文本，模型就会崩溃成一种“向评分量表中心压缩”的幻觉（或者说是训练集过拟合）。<p>X 上的某人让我试着复现这个问题，我实际上在他们的 Gemini Chat 上第一次尝试就成功了：<p>https://x.com/XCSme/status/2057613611959279988<p>我不知道该怎么看待这个问题（或者大多数 SOTA 模型）。它们在编码和工具使用方面变得更聪明了，但在其他方面却变得更笨了...
查看原文
I thought I was going crazy, trying to use Gemini 3.5 Flash to rate some answers, but it kept giving 7 instead of 10 for correct answers.<p>Apparently once you add a "Grading criteria" text, the model collapses into a "compressed toward the center of the scale" hallucination (or training set overfitting).<p>Someone on X asked me to try to reproduce it, and I actually got it on the first try on their Gemini Chat:<p>https://x.com/XCSme/status/2057613611959279988<p>I am not sure what to make of this (or most SOTA) models. They got a lot smarter with coding and tool usage, but a lot dumber in other ways...