引用avereveard · 2023-08
“Further reads”
列入延伸阅读:语言模型能否分辨自己在编造参考文献
靠一致性提问,让模型自己识别出编造的参考文献
论文把幻觉的书籍和文章引用当作研究语言模型幻觉的“模式生物”:不查外部资源,只向模型追问作者等细节,做一致性检查,就能识别出编造的引用。结果显示,包括 GPT-4 在内的模型对编造的引用常给出前后不一致的作者列表,而对真实引用往往能稳定答对,可以说模型“知道”自己在编造。
HN 上的人用它来说明“幻觉的答案每次问都不一样,而事实性的答案更一致”,并支持模型内部其实有足够信息判断自己是否在出错。
“Further reads”
“In particular, we find that LMs often hallucinate differing authors of hallucinated references when queried in independent sessions, while consistently identify authors of real references.”
“Apparently research suggests that when LLMs hallucinate answers, their hallucination is likely to be a different one every time, where as factual answers will tend to be consistent.”
“There are various results that suggest that LLMs do internally have everything they'd need to know that they're hallucinating/wrong:”