On a 200-item synthetic probe whose answers depend on preserving or inspecting the model's own control-literal text, nameless tokenization achieved 59.9% overall accuracy versus 8.5% for standard tokenization, 2.2% for stripping, 23.0% for masking, and 24.1% for escaping, while also leading each of the four independent probe columns.

来源：fixed-paper:PDF pp. 3–4 and 8, Table 2, Section 4, and Appendix B