@hauschke @SemAntiKast Hmm, now I'm not sure what discussion you're referring to.
We've mainly tried LLMs for metadata extraction, not classification as such. I gave a talk about it at #SWIB24 with Pierre Beauguitte from Norway (and also a lightning talk at #SWIB23 ).
Code for the experiments can be found here: https://github.com/NatLibFi/FinGreyLit
Most experiments involve fine-tuning smallish LLMs. Haven't tried reasoning models yet.