Mapping Token Transitions via Repeated Token Input Analysis
September 28, 2026
This research analyzes the argmax probability of a model's response to its own repeated token input to identify black-box model characteristics. The study identifies fixed points in the vocabulary and maps how non-fixed point tokens transition, providing a method for model identification.
HOW THIS AFFECTS YOU
●
researcherYou can use this token-mapping technique to probe the internal behavior and identity of black-box models.