The most recent cross encoders on Huggingface are from like 2 years ago.
It makes no sense to me though because it would seem that with LLMs being all the rage, getting good rerankings are especially important nowadays because of the LLM token limit and cost per token.
With vector retrieval, you can go from millions of docs to a decent set of top 50, but you can't pass 50 docs to an lLM so you'd probably want to rerank those 50 and get a really good top 3 docs to feed to the LLM.
Yet there's no advancements in cross-encoders at all??? What is going on, I've been looking for days and I cant find anything.
(just to be clear, I'm talking about models that take query + passage together to allow attention calculations across the two, and then outputs some similarity score)