2026-08-06 16:37 UTC
@kfort@social.sciences.re Probably yes. But inclusion in pre-training is not the same as memorization, and the paper explicitly discusses this limitation.
Replies (1)
-
@kfort@social.sciences.re 2026-08-06 16:41
@serhii@mstdn.science Hum. So the data from the papers were used to pre-train the LLM. Of course, it does not imply that the LLM will regurgitate the content, but... in reality, esp. with big models, it does. I seem to remember this is a good paper about this (but there are plenty of others): https://aclanthology.org/2021.eacl-main.113/