A Unified Framework of Extractability in Large Language Models
Santiago Morón Santos, Matías Carrasco, Sergio YovineThis paper proposes a formal framework for analyzing extractability in language models based on the minimum number of attempts required to extract a set of sequences with a given confidence. It unifies previously proposed notions and mitigates the theoretical fragmentation found in the literature. Moreover, it covers both verbatim and approximate extractability and can be applied to evaluate model performance as well as memorization. Our work brings up the importance of string-generation parameters like masking, length and temperature scaling, together with evaluation parameters like string similarity. The experiments carried out on several realistic scenarios show the effect of these parameters on extractability measured in state-of-the-art language models.