Authors' Reply: Citation Accuracy Challenges Posed by Large Language Models

"We also propose another solution aimed at fundamentally reducing citation errors: the development of “Reference-Accurate” academic LLM by major global publishers. Leading journals could develop their own specialized LLM, trained exclusively on rigorously verified academic literature from robust databases. This targeted training would ensure that every generated reference is accurate and directly traceable to published work. Ideally, these publisher-backed LLMs would be made freely available to promote open science."

I don't think it works like that. RAGs might provide something, yes, but they're not applicable due to paywalls. Funnily enough, some implementations in this space (https://news.ycombinator.com/item?id=44274487) do not even give access to the papers so you're basically point-blank trusting a LLM.