Tag
This paper investigates black-box model identification by using a repeated token as degenerate input to map token destinations, showing that non-fixed-point tokens can accurately distinguish model families and measuring robustness across quantization levels.