Tag
Yohei Nakajima presents a method to read typed visual judgements from a frozen open vision-language model using logits, achieving similar accuracy to hosted models with reduced time and GPU cost.