“a multi-step method that iteratively corrects and re-embeds text is able to recover 92% of 32-token text inputs exactly”.