Hallucination Is Inevitable: An Innate Limitation of Large Language Models (opens in new tab)

(arxiv.org)

3 pointsPerryCox2y ago2 comments

2 comments

This is pretty obvious no? LLMs are basically a lossy compression of their dataset. Being lossy, they will necessarily have error (hallucinations). Furthermore, the underlying data is human approximation of truth. Therefore it will have error as a well. Shall we publish a paper on the linked list?

rhelz2y ago

Hmmm....couldn't you argue it the other way as well? I mean, a hallucination isn't just an incomplete fact present in the input set--it is a claim which wasn't in the input set.

That's not what we'd expect from a lossy compressor; we would expect incomplete or missing answers, but we wouldn't expect it to give us claims which were not in the input set.

j / k navigate · click thread line to collapse

2 comments

rapatel02y ago

rhelz2y ago

Hmmm....couldn't you argue it the other way as well? I mean, a hallucination isn't just an incomplete fact present in the input set--it is a claim which wasn't in the input set.

That's not what we'd expect from a lossy compressor; we would expect incomplete or missing answers, but we wouldn't expect it to give us claims which were not in the input set.

j / k navigate · click thread line to collapse