> Giving few-shot examples to the generator didn't work.
> Using an LLM-judge (with no training) didn't work.
> Using an embedding + KNN-classifier (lots of training data) worked.
I don't know why they didn't try fine-tuning the LLM-judge, or at least give it some few-shot examples.
But it shows that embeddings can make very simple classifiers work well.
> Giving few-shot examples to the generator didn't work.
> Using an LLM-judge (with no training) didn't work.
> Using an embedding + KNN-classifier (lots of training data) worked.
I don't know why they didn't try fine-tuning the LLM-judge, or at least give it some few-shot examples.
But it shows that embeddings can make very simple classifiers work well.