引用lambdaviking · 2023-05
“As some evidence of this, we were able to use these results to find simple problems that ChatGPT and GPT4 get chance performance on.”
引用它:用这些结果找出了 ChatGPT 和 GPT-4 表现仅相当于随机的简单问题
模型为圆早先的错误而继续出错,即使它能识别这些错
作者构造了三个问答数据集,ChatGPT 和 GPT-4 常先给出错误答案,再给出含错误论断的解释。关键发现是,这两个模型能分别识别出自己 67% 和 87% 的错误。论文把这种现象称为“幻觉滚雪球”:模型过度坚持早先的错误,由此产生本来不会出现的更多错误。
HN 上的人把它当作“早期错误会在后续生成中累积放大”的早期论文来引用,用来解释模型一旦说错就越说越错。
“As some evidence of this, we were able to use these results to find simple problems that ChatGPT and GPT4 get chance performance on.”
“How Language Model Hallucinations Can Snowball] (Zhang et al., 2023)”
“they also reference the original snowballing paper”
“Off the top of my head - here's one of the earlier papers on consistency-induced hallucinations:”