Tag
This paper explores the feasibility of reconstructing the tokenizer of a closed-source large language model using only two oracle queries from its chat API, potentially revealing internal representation details.