Nameless tokenization removes the added-token text lookup and sentence-marker post-processor from a reloaded content encoder, renders chat-template control literals separately as identifiers, and encodes all remaining template text and message content in maximal runs. Under the paper's text-interface threat model, this makes it impossible for attacker-chosen text to emit a template control identifier while leaving every message character unchanged. · CiteArk
Nameless tokenization removes the added-token text lookup and sentence-marker post-processor from a reloaded content encoder, renders chat-template control literals separately as identifiers, and encodes all remaining template text and message content in maximal runs. Under the paper's text-interface threat model, this makes it impossible for attacker-chosen text to emit a template control identifier while leaving every message character unchanged.
