vertical
(horizontal)
independent parts
are these just pretokens?
Can we feasibly (uniformly) sample from the set of all tokenizations of a string?
Yes: https://aclanthology.org/2025.acl-long.1180/
one work
another nice one is Adversarial Tokenization: https://aclanthology.org/2025.acl-long.1012/
default
known in the trade as canonical.
(horizontal)